DE102022210424A1 - Self-localization of a rail vehicle - Google Patents

Abstract

A method for determining the probability distribution (p(P)) of a pose (P) of a rail vehicle (1) is described. In the method according to the invention, a formal structure of a system state (x), which represents a pose (P) of the rail vehicle (1), and a motion model (BM) of the rail vehicle (1), which represents a probability distribution of a pose change (DP) as a function of time, are defined, wherein a possible pose (P) of the rail vehicle (1) is restricted on the basis of the course of rail routes in track map data (GKD), but a partial compensation of the restriction takes place on the basis of a model-based defined possible deviation of the pose (P) of the rail vehicle (1) from the restricted possible pose (P). In addition, sensor models (SM) are defined for characterizing the measurement behavior of sensors of the rail vehicle (1), which are used for determining the probability distribution (p(P)) of the pose (P) of the rail vehicle (1) on the basis of sensor data (SD). Furthermore, a Bayes filter (BF) is determined on the basis of the system state (x), the motion model (BM) and the sensor models (SM). In addition, a probability distribution (p(x0)) of an initial state (x0) of the rail vehicle (1) is configured. During a journey of the rail vehicle, sensor data (SD) is recorded by the sensors to determine the pose (P) of the rail vehicle (1) and the probability distribution (p(P)) of the pose (P) of the rail vehicle (1) is determined by applying the Bayes filter (BF) to the recorded sensor data (SD) and to the probability distribution (p(x0)) of the initial system state (x0) of the rail vehicle (1) as input variables. A self-localization device (30) is also described. A rail vehicle (1) is also described.

Description

The invention relates to a method for determining the probability distribution of a pose of a rail vehicle. The invention also relates to a self-localization device. The invention also relates to a rail vehicle.

Knowledge of the pose of a rail vehicle is required for many different applications in the field of rail transport. When we talk about the pose of a rail vehicle, this value includes both the position of the rail vehicle and its orientation in space. For example, passenger information systems can provide passengers with information related to the current position of the rail vehicle. This data can also be used to plan train operations. For example, based on the determined position of a train, a train delay can be predicted and displayed, or block distances between trains can be reduced. In addition, necessary maintenance can be predicted. In this context, information regarding the geometry or course of a track section can be used to evaluate vibration measurements. Pose data can also be used to implement automated signal detection by predicting the position of a signal in a camera image and thus support an ADAS system (ADAS = Advanced Driver Assistance System). Pose data can also be used for obstacle detection and a reaction to such an obstacle. On this basis, a trajectory can be determined in order to classify detected objects as critical or non-critical.

The accuracy requirements for a localization solution vary considerably depending on the application.

For example, passenger information systems only require precision in the meter range and the orientation of the rail vehicle is unimportant. For obstacle detection, on the other hand, the position and orientation of the rail vehicle must be known with high accuracy in order to detect and precisely localize obstacles at large distances, for example a few 100 meters to 1 km away.

Conventional methods for locating rail vehicles can be based on a naive “map-snapping” algorithm. A single piece of sensor information, typically a GNSS position (GNSS = Global Navigation Satellite System), is compared with a rail map. In the rail map, the closest point to this GNSS position on a rail track is selected and classified as the most likely position of the rail vehicle and this position is output to a user as the position of the rail vehicle.

A second approach or a whole group of approaches can be classified as a solution approach based on the application of a Bayesian filter.

In a Bayesian filter, a system state is estimated and tracked over time, such that subsequent consideration of additional sensor measurement data leads to an improved estimate. Bayesian filters for rail vehicle localization comprise a motion model that predicts the likely motion of a rail vehicle between two points in time, and a set of sensor models that update the likely state of the estimated pose whenever sensor data becomes available.

eines physikalischen Systems über die Zeit zu verfolgen.A Bayesian filter is used to determine an n-dimensional system state $$x=\left[\begin{array}{c}{x}^1\\ ⋮\\ x^n\end{array}\right]$$

of a physical system over time.

As a rule, it is sensible not to define the system state in the Bayesian filter solely by the variables that directly indicate the pose of a rail vehicle. Instead, the estimation can be significantly improved if additional variables are tracked over time. For example, an estimate of the current speed of the rail vehicle in the motion model can provide a significantly better prediction of the system state than based solely on the last estimated position or pose of the rail vehicle. Other values that can represent part of the system state include the bias values for individual sensors, in particular inertial sensors, or the clock error in a GNSS receiver. The motion model therefore includes the physical movement of the rail vehicle, but it also describes the temporal development of other physical quantities that are part of the system state. The term motion model therefore does not only refer to the physical movement of the rail vehicle. tool, but describes more generally the temporal development of the system state.

The estimate of state, also called system state or motion state, is represented in probabilistic form, for example as probability density, also called distribution density, for continuous state variables. If the state also contains discrete variables, a discrete probability distribution or a combination of discrete distribution and distribution density is required. The following equations assume continuous state variables and thus distribution densities in order to simplify the notation.

The state estimate is based on an initial estimate of the state, which usually represents a very high uncertainty, since the system state can only be estimated more precisely as measurement data is received. The current state estimate is initially determined by a motion model. Such a motion model is represented by the conditional probability density: $$p\left(x_k|x_{k-1}\right).$$

This expression describes the conditional probability of the state x at time k depending on the state x at time k-1. In addition, the estimate is updated at discrete times according to equation (2) as a result of measurements or sensor data z, which are represented by one or more sensor models. Such a sensor model is represented by the probability distribution: $$p\left(z_k|x_k\right).$$

The sensor model describes a conditional probability distribution of sensor measurement data z at time k depending on the state x at time k.

The update of the state estimate by the movement model and the update by the sensor models typically take place alternately, so that an update by the movement model takes place before each update with sensor data. Bayes' theorem provides the equation to determine the estimate of the updated motion data from the sensor measurement.

die Aktualisierung durch das Sensormodell gemäß: $$p\left(\left(x_k|z_k\right)\right)=\mathrm{\eta }\cdot p\left(z_k|x_k\right)\cdot \tilde{p}\left(x_k\right),$$

wobei η eine Normalisierungskonstante ist, die zwar allgemein schwer zu berechnen ist, jedoch in der Regel nicht benötigt wird.The update by the movement model is carried out according to: $$\tilde{p}\left(x_k\right)={\displaystyle \int p\left(x_k|x_{k-1}\right)d{x}_{k-1},}$$

the update by the sensor model according to: $$p\left(\left(x_k|z_k\right)\right)=\mathrm{\eta }\cdot p\left(z_k|x_k\right)\cdot \tilde{p}\left(x_k\right),$$

where η is a normalization constant that is generally difficult to calculate but is usually not needed.

The algorithm described by the equations for updating by motion model and sensor model is also called Markov localization due to the recursive state calculation from the previous time step in connection with localization. This Markov localization is described, for example, in Sebastian Thrun et al. "Probabilistic Robotics (Intelligent Robotics and Autonomous Agents)". The MIT Press, 2005.

However, the exact calculation of the probability densities or probability distributions of a system state at time k is not possible as a result of the aforementioned equations for general formulations of the motion and sensor models. Under certain simplifying conditions for the motion model, the sensor models and the probability distribution or probability density of the system state, special filters can be used that allow efficient and precise calculation of the update steps. The classic Kalman filter can be used for linear Gaussian systems. For non-linear update models, the Extended Kalman Filter, the Unscented Kalman Filter or the Error-State Extended Kalman Filter can be used under certain restrictions (such as quasi-Gaussian distributions). For systems with significantly non-Gaussian probability distributions, however, multi-hypothesis filters can be used under certain conditions.

If the probability density is distributed over individual significant peaks, which ideally show, for example, approximately Gaussian behavior, it is possible to track the individual peaks as separate hypotheses, in particular by means of powerful filters such as the Extended Kalman filter and the Unscented Kalman filter. In parallel to the parameters of the individual hypotheses (for example, mean and covariance), the weight of the individual hypotheses is then tracked, from which the function for the entire distribution follows through weighted addition of the individual hypotheses.

The maximum flexibility towards nonlinear update equations and non- Gaussian distributions are given by the particle filter, also known as the sequential Monte Carlo filter. In addition, the Rao-Blackwell particle filter offers the possibility of combining the particle filter with, for example, the Kalman filter.

• Zum einen existieren kommerzielle Lokalisierungslösungen, welche eine zentimetergenaue Position und eine geforderte Genauigkeit der Orientierung liefern können. Diese Verfahren basieren gewöhnlich auf Echtzeitkinematiken (RTK = Real Time Kinematics) und gekoppelter Inertialnavigation. Diese Genauigkeiten werden allerdings typischerweise nur unter idealen Umweltbedingungen erreicht, bei denen Satellitendaten verfügbar sind. Außerdem sind die leistungsstärksten dieser Systeme besonders teuer aufgrund der Nutzung hochpräziser Trägheitsmesstechnologie. Bei diesen kommerziellen Lösungen wird typischerweise die Fahrzeugpose in einem Lösungsraum mit vollen 6 Freiheitsgraden geschätzt, wobei Einschränkungen, die sich für Schienenfahrzeuge anhand von Gleis-Kartendaten ergeben, nicht berücksichtigt werden. In einigen Fällen wird eine teilweise Nutzung von Gleis-Kartendaten untersucht, um die Lokalisierung auf eine breitere Datenbasis zu stellen, wobei die spezielle Dynamik in Folge des Schienenverlaufs berücksichtigt wird. Eine solche Vorgehensweise ist in C. Reimer et al. „INS/GNSS/Odometer Data Fusion in Railway Applications“. In Symposium Inertial Sensors & Systems 2016. beschrieben.

Localization methods for rail vehicles that use Bayesian filters can be divided into two major classes:

• On the one hand, there are commercial localization solutions that can provide a centimeter-accurate position and the required orientation accuracy. These methods are usually based on real-time kinematics (RTK) and coupled inertial navigation. However, these accuracies are typically only achieved under ideal environmental conditions where satellite data is available. Additionally, the most powerful of these systems are particularly expensive due to the use of high-precision inertial measurement technology. These commercial solutions typically estimate vehicle pose in a solution space with a full 6 degrees of freedom, ignoring limitations that arise for rolling stock from track map data. In some cases, partial use of track map data is being investigated to provide localization on a broader database, taking into account the specific dynamics resulting from the track's path. Such an approach is described in C. Reimer et al. “INS/GNSS/Odometer Data Fusion in Railway Applications”. Described in Symposium Inertial Sensors & Systems 2016.

A possible representation of a system state x for a Bayesian filter that does not use track maps would be: $$x=\left[\begin{array}{c}lOn\\ lat\\ alt\\ \phi \\ \theta \\ \psi \\ v_x\\ v_y\\ v_z\\ b_{ax}\\ b_{ay}\\ b_{az}\\ t_{err}\end{array}\right].$$

Here, lon, lat, alt are the so-called WGS84 coordinates (The World Geodetic System 1984 is a geodetic reference system as a uniform basis for position information on Earth and in near-Earth space), φ, θ, ψ are the Euler angles that indicate the orientation of the rail vehicle, v _x , v _y , v _z are the linear speed values in three dimensions, b _ax , b _ay , b _{az are} the bias values for a 3D accelerometer and t _{err is} a GNSS clock error. Part of the state x are therefore also 6 coordinates lon, lat, alt, φ, θ, ψ, which describe a pose P of the rail vehicle.

Sensor bias refers to quasi-static errors in measuring devices that only change very slowly, for example depending on the ambient temperature. Failure to take such errors into account is particularly problematic when the measured values are integrated over time, for example when determining speed and position by integrating acceleration, since constant components of the measurement error add up in the state estimate (drift). At the same time, it is possible to minimize the influence of the quasi-static errors by using other sensor modalities to estimate the bias values online as part of the system state. Estimating bias values in acceleration sensors is possible, for example, by measuring the absolute position using drift-free sensors such as GNSS.

As an example of a motion model used in the literature for an approach without the use of track maps, reference is made here to Sebastian Thrun et al. "Probabilistic Robotics (Intelligent Robotics and Autonomous Agents)". The MIT Press, 2005. This describes a motion model for a mobile robot in the plane. The "Velocity Motion Model" calculates the resulting probability distribution at time k for the state variables x, y, ψ, i.e. the 2D position and the yaw angle, in the plane. In this example, the translational and rotational speeds are considered as control variables, i.e. as known input variables for the state update. State models for commercial localization solutions naturally consider higher-dimensional state representations and more complex motion models.

The implementation of a suitable Bayes filter for such a high-dimensional system state is usually realized with a so-called extended Kalman filter or an error state Kalman filter, which can handle high system dimensions and can map non-linear relationships between motion models and sensor models.

Another conventional approach to localizing rail vehicles uses the known track layout of the tracks on which a rail vehicle to be localized is currently located to restrict the solution space, since the rail vehicle must be located somewhere on the tracks. Such an approach is described, for example, in O. Heirich. “Localization of Trains and Mapping of Railway Tracks”. PhD Thesis, 2020.

In this way, the localization problem is usually traced back to a 1D localization problem, whereby the Bayesian filter only needs to answer the question of what distance the rail vehicle has currently traveled. Several hypotheses can also be included, which take into account that the rail vehicle drives over a switch with an unknown status. However, this approach may not achieve a particularly high level of accuracy, since this model generally cannot accurately represent the real state of a rail vehicle, since it does not allow the pose of the rail vehicle to deviate from a pose determined by the course of a rail route.

A possible definition of a system state x of a rail vehicle using a track map is represented by: $$x=\left[\begin{array}{c}id\\ s\\ d\\ v\end{array}\right].$$

id, s and d represent the topological pose of the rail vehicle and v the speed of the rail vehicle along the track. Specifically, id represents the track segment ID, s represents the offset on the track segment, and d represents the direction on the track segment. In such a state representation, the geographical pose of the rail vehicle is clearly specified by the topological pose. Deviations, for example, of the yaw angle or the roll angle compared to the track geometry, as can occur due to the elastic suspension of the car body or due to the distance between the front and rear axles in curves, are not recorded by such a state model or such a structure of a system state.

$${\text{d}}_{\text{k}}\to {\text{d}}_{\text{k}-1}$$

$${\text{s}}_{\text{k}}\to {\text{s}}_{\text{k}-1}+\text{v}\Delta \text{t}$$

$${\text{v}}_{\text{k}}\to {\text{v}}_{\text{k}-1}+{\mathrm{\sigma }}_{\text{v}}\text{w},$$

wobei Δt die Zeitdifferenz zwischen den Zeitpunkten k-1 und k beschreibt und σ_vw einen gaussförmigen unabhängigen Rauschprozess mit Intensität σ_v beschreibt (w wird aus einer Einheits-Gaußverteilung mit einem Mittelwert von 0 und einer Standardabweichung von 1 generiert). Auf einer komplexeren Gleiskarte muss das Bewegungsmodell außerdem abbilden, wie die unterschiedlichen Gleissegmente miteinander verbunden sind, so dass bei Überschreiten eines Segmentendes oder -anfangs die Transition auf benachbarte Segmente erfolgt. Eine konkrete Implementierung dieser Vorgehensweise ist zum Beispiel in O. Heirich. „Localization of Trains and Mapping of Railway Tracks". PhD Thesis, 2020. beschrieben.For the sake of illustration, a simplified track-based motion model is described below. To simplify, it is assumed that the track map consists of a single track, so that, for example, the track segment ID id = 1 can always be assumed. In addition, the original orientation d ₀ of the rail vehicle on the track is known, for example d ₀ = “forward”. The movement model defines the following transition for the state variables: $${\text{Id}}_{\text{k}}\to {\text{id}}_{\text{k}-1}$$

$${\text{d}}_{\text{k}}\to {\text{d}}_{\text{k}-1}$$

$${\text{s}}_{\text{k}}\to {\text{s}}_{\text{k}-1}+\text{v}\Delta \text{t}$$

$${\text{v}}_{\text{k}}\to {\text{v}}_{\text{k}-1}+{\mathrm{\sigma }}_{\text{v}}\text{w},$$

where Δt describes the time difference between times k-1 and k and σ _v w describes a Gaussian independent noise process with intensity σ _v (w is generated from a unit Gaussian distribution with a mean of 0 and a standard deviation of 1). On a more complex track map, the movement model must also show how the different track segments are connected to each other, so that when the end or beginning of a segment is exceeded, the transition to neighboring segments occurs. A concrete implementation of this approach is, for example, in O. Heirich. "Localization of Trains and Mapping of Railway Tracks". PhD Thesis, 2020.

In the above example with only one track segment and known orientation of the vehicle, id and d can be removed from the system state and the motion model, since they are always unchanged and known and the transition probability in equation (2) for s and v follows directly from equation (10) and equation (11).

As an example of a sensor model for both track card-based and track card-free approaches, a Gaussian GNSS position sensor model will be described here. In such a GNSS position sensor model, the measurement is disturbed by a Gaussian error term. The sensor model p(z ^GNSS |x) is described by a simple multidimensional Gaussian distribution N: $$Z^{GNSS}\sim N\left(\mu ,\Sigma \right).$$

The mean value µ is represented by the longitude, latitude and altitude above zero, which result from the current system state.

• - für den Systemzustand ohne Gleiskarte direkt aus dem Längengrad lon, dem Breitengrad lat, und der Höhe alt im Systemzustand,
• - für den Systemzustand mit Gleiskarte indirekt aus den topologischen Parametern Gleissegment-ID id, Offset s auf dem Gleissegment (die Orientierung d auf dem Gleis spielt für die Position keine Rolle), wobei die Geometrieinformation aus der Gleiskarte bezogen werden muss, also eine Funktion f: id, s -> lon, lat, alt (13).

These values follow:

• - for the system state without track map directly from the longitude lon, the latitude lat, and the altitude alt in the system state,
• - for the system state with track map indirectly from the topological parameters track segment ID id, offset s on the track segment (the orientation d on the track plays no role for the position), whereby the geometry information must be obtained from the track map, i.e. a function f: id, s -> lon, lat, alt (13).

The covariance matrix Σ can be assumed to be constant in a simple model, but it is also possible to take into account the actual uncertainty of a GNSS measurement, for example due to the number and constellation of available satellites.

As already mentioned, the approaches described in detail above bring with them certain, partially complementary, limitations when it comes to the accuracy of pose determination.

The task is therefore to enable self-localization of a rail vehicle that is more precise, reliable and robust compared to the state of the art.

This object is achieved by a method for determining a probability distribution of a pose of a rail vehicle according to patent claim 1, a self-localization device according to patent claim 11 and a rail vehicle according to patent claim 12.

In the method according to the invention for determining a probability distribution of a pose of a rail vehicle, in step i) a formal structure of a system state, which represents a pose of the rail vehicle, and a motion model of the rail vehicle, which represents a probability distribution of a pose change as a function of time, are determined or defined. A "formal structure" is used here to describe that at this point no concrete system state with a concrete pose is determined, but rather a mathematical formula structure or the mathematical form of the system state, as described for example further below in equation (14), is formulated or defined as a model. As already explained, the system state can include further information about the state of the rail vehicle in addition to the pose, such as the speed of the rail vehicle. In this case, a possible pose of the rail vehicle is restricted on the basis of the course of rail routes in track map data, but the restriction is partially compensated on the basis of a model-based defined possible deviation of the pose of the rail vehicle from the restricted possible pose. In particular, the movement model can be adapted to concrete scenarios based on specific parameter values.

In addition, in step ii), sensor models are defined for characterizing the measurement behavior of sensors of the rail vehicle, which are used to determine the probability distribution of the pose of the rail vehicle based on sensor data. In practice, what the sensor models look like is ultimately determined by which sensors are available in a particular application.

Furthermore, the sensor models can be adapted to specific scenarios based on specific parameter values, similar to the motion model.

Based on the system state, the movement model and the sensor model, a Bayesian filter is determined in step iii). As already explained, a Bayes filter includes a recursive probabilistic method for estimating probability distributions of unobserved system states based on observed measured values. The Bayes filter is later used to determine a new or current system state and a corresponding pose of the rail vehicle based on a last determined system state and based on recorded sensor data.

In addition, in step iv) a probability distribution of an initial system state of the rail vehicle is configured. Such an initial system state can be known, for example, based on external information. For example, it is known with predefined accuracy from which position in the rail network or with which pose the rail vehicle starts.

Furthermore, in step v), sensor data is recorded by the sensors to determine the pose of the rail vehicle.

Sensor data refers to measurement data that is suitable for determining poses. Typical measurement data are GNSS data, odometry data, sensor data for scanning the environment with which landmarks can be detected, etc.

Finally, in step vi), a probability distribution of a system state and thus also a probability distribution of the pose of the rail vehicle is determined by applying the Bayes filter to the recorded sensor data and to the probability distribution of the initial system state of the rail vehicle as input variables.

Preferably during a journey of the rail vehicle, steps v) and vi) are then repeated several times, with an updated pose of the rail vehicle or its probability distribution being determined in each case.

For this purpose, in step vi), instead of the probability distribution of the initial system state of the rail vehicle, a probability distribution of a system state of the rail vehicle last determined in connection with the last determined probability distribution of the pose of the rail vehicle is used as an input variable for the Bayesian filter.

In other words, the last estimated system state after each further iteration or repetition of steps v) and vi) includes the system state last determined when the Bayesian filter was used to determine the pose. The steps for acquiring sensor data and for determining the probability distribution of the pose of the rail vehicle are therefore carried out or iterated repeatedly during a journey in order to create a new current system state and a new one from the last determined pose or the last determined system state comprising this pose to determine the current pose of the rail vehicle.

The knowledge of the likely pose of the rail vehicle can be used for pose-dependent actions, reactions, especially applications described in the introduction.

Based on the sensor data and on the basis of track map data, preferably high-precision track map data, which include or represent a section of route on which the rail vehicle is currently located, a limited solution space is clearly stated, which includes the sought pose, but deviations of the pose from one Pose determined based on distance alone is allowed to be determined. The restricted solution space results from the knowledge of the nominal trajectory given by the track geometry on which the rail vehicle is located, as well as from the knowledge of possible deviations of the pose of the rail vehicle from this trajectory. The trajectory is determined based on the course of a rail section or route section, which is assigned to the rail vehicle on the basis of the preferably highly accurate map data as well as the sensor data and older pose data. To restrict the solution space, it can initially simply be assumed that the deviations from a nominal behavior of the rail vehicle pose permitted in the restricted solution space are “small” compared to the amounts of the state variables that would be possible without the restriction caused by the track data. Highly accurate track map data is displayed with a predetermined accuracy. The accuracy required depends on the applications for which self-location is to be used. Highly accurate track map data with a maximum relative deviation of 5 cm and a maximum absolute deviation of 20 cm are preferably displayed. The relative deviation is particularly preferably 5 cm and the maximum absolute deviation is only 10 cm. With greater accuracy of the track map, the accuracy that can be achieved when determining the pose of the rail vehicle also increases. It should be noted at this point that with the method according to the invention, relatively imprecise track maps can also be used, which cause problems in the prior art, since according to the invention, deviations in the pose of the rail vehicle compared to the track map are permitted.

A motion model defines the transition of the state variables of a system state x between the times k-1 and k. Examples of such motion models are described in equations (10) and (11) or in the “Velocity Motion Model” in Sebastian Thrun et al. “Probabilistic Robotics (Intelligent Robotics and Autonomous Agents)”. The MIT Press, 2005.

Illustrated. A sensor model defines a relationship between a system state and the probability distribution/distribution density of a specific measured value or sensor measured value. An example of such a sensor model is illustrated in equations (12) and (13). The definition of a system state and a motion model is advantageously simplified due to the assumption of a restricted solution space compared to a conventional approach without such a restriction, as will be illustrated in detail later. On the other hand, an increased precision of an estimate of a system state and thus of a pose of a rail vehicle is made possible because the method allows deviations of the position and pose from a trajectory of a section of track on which the rail vehicle is currently located.

The pose of the rail vehicle is preferably determined by applying a Bayesian filter to a combination of a motion model and a set of sensor models. In other words, an approximate pose of the rail vehicle is determined on the one hand based on preferably highly accurate track map data. On the other hand, a reduced solution space is determined on the basis of the approximate pose, in which an exact pose of the rail vehicle is likely to be found. The Structure of a filter algorithm is considerably reduced due to the restriction of the solution space, so that the computational effort is reduced compared to a generic approach, especially compared to track-mapless approaches.

However, unlike a conventional track map-based approach, deviations of the pose of the rail vehicle from a "nominal" behavior of the rail vehicle in the motion model or from the approximate pose determined based on the track map in the reduced solution space are determined in order to determine more precise values for a pose of the rail vehicle.

The system state x of a rail vehicle can now be described as follows: $$x=\left[\begin{array}{c}id\\ s\\ d\\ v\\ \Delta \phi \\ \Delta \theta \\ \Delta \psi \\ \Delta y\\ \Delta z\end{array}\right].$$

Here, id, s, d represent the topological pose of the rail vehicle, v the speed along the track, Δφ, Δθ, Δψ, Δy, Δz the pose deviation, i.e. the deviation of the orientation and the translation in transverse and vertical directions.

However, the system status can still be represented in a simplified manner. This is because certain properties of the physical system can be taken into account in order to reduce the dimension of the system state. For example, the deviations Δψ of the rail vehicle pose compared to the track geometry around the yaw axis or the deviations Δφ compared to the track geometry around the roll axis are considerably more significant than the deviations Δθ around the pitch axis. The translational deviations Δy, Δz can also be neglected in many applications. In this way a simplified system state of the form is obtained: $$x=\left[\begin{array}{c}id\\ s\\ d\\ v\\ \Delta \phi \\ \Delta \psi \end{array}\right].$$

Δφ indicates the roll angle deviation and Δψ indicates the yaw angle deviation.

The inclusion of track maps advantageously greatly reduces the solution space for determining the system status. In contrast to conventional solutions with track maps, deviations in the pose of the rail vehicles from the track geometry are also permitted and tracked.

Due to the restriction to an area around the track, pose deviations can only be very small, which means that a linear representation with a Gaussian distribution is suitable for calculating the deviations. The small deviations also make it possible to consider different degrees of freedom in a decoupled manner.

In the movement model according to the invention, in addition to the transition of the topological state, a transition for the pose deviation or the deviation of the orientation is also defined. In the case that only a deviation in the yaw angle is considered, this is then, for example $$\Delta {\mathrm{\psi }}_{\text{k}}\to \Delta {\mathrm{\psi }}_{\text{k}-1}+{\mathrm{\sigma }}_{\mathrm{\psi }}\text{w}.$$

The term σ _ψ w in equation (16) expresses, similarly to the speed in equation (11), which changes in the yaw angle deviation Δψ _k are to be expected between two times k - 1 and k. The pose deviation update can also include more complex representations if necessary, for example to reflect inertial effects. However, linear behavior can still be assumed here.

For the system status with track map plus pose deviation, the sensor model can also be determined from the track map.

For a sensor model for position measurement, the geocoordinates from the track map must be adjusted by the position deviations Δy and Δz. The following function f can then be defined: $$\text{f}:\text{id},\text{s},\Delta \text{y},\Delta \text{e.g}\to \text{lon},\text{lat},\text{old}.$$

The self-localization device according to the invention has a filter determination unit for determining a Bayesian filter by defining a formal structure of a system state, which represents a pose of a rail vehicle, and a movement model of the rail vehicle, which represents a change in pose of the rail vehicle as a function of time represents, on. A possible pose of the rail vehicle is restricted based on the course of rail routes in track map data, but the restriction is partially compensated for based on a model-based possible deviation of the pose of the rail vehicle from the restricted possible pose.

The filter determination device is also designed to define sensor models for characterizing the measurement behavior of sensors of the rail vehicle, which are used to determine the probability distribution of the pose of the rail vehicle based on sensor data.

In addition, the filter determination device is configured to determine a Bayesian filter based on the system state, the motion model and the sensor model.

Part of the self-localization device according to the invention is also a configuration unit for configuring a probability distribution of an initial state of the rail vehicle.

Furthermore, the self-localization device according to the invention comprises a sensor data interface for acquiring sensor data from sensors of the rail vehicle to determine the pose of the rail vehicle and a pose determination unit for determining a probability distribution of a current system state and thus also the probability distribution of a pose of the rail vehicle by applying the Bayes filter to the recorded sensor data on the probability distribution of the initial system state of the rail vehicle as input variables.

Preferably, the pose determination unit is set up to use a probability distribution of a last estimated system state of the rail vehicle as an input variable for the Bayesian filter instead of the probability distribution of the initial system state in a repeated pose determination. After each further iteration, this last estimated system state includes the system state of the rail vehicle that was last determined when the Bayesian filter was used to determine the pose.

Therefore, the pose determination unit is designed to use a probability distribution of a system state of the rail vehicle last determined in connection with the last determined probability distribution of the pose of the rail vehicle as the input variable for the Bayes filter during a repeated pose determination, instead of the probability distribution of the initial system state of the rail vehicle.

The self-localization device according to the invention shares the advantages of the method according to the invention for determining a probability distribution of a pose of a rail vehicle.

The rail vehicle according to the invention has a sensor unit for recording sensor data for determining a pose of the rail vehicle. In addition, the rail vehicle according to the invention comprises the localization device according to the invention, which is set up to determine the pose of the rail vehicle on the basis of the recorded sensor data. Optionally, the rail vehicle according to the invention comprises a device for carrying out an action based on knowledge of the pose of the rail vehicle, such as a control device for controlling a driving behavior of the rail vehicle depending on a pose determined by the localization device. The rail vehicle according to the invention shares the advantages of the self-localization device according to the invention.

A majority of the aforementioned components of the self-localization device can be implemented entirely or partially in the form of software modules in a processor of a corresponding computing system, for example by a control unit of a rail vehicle. A largely software-based implementation has the advantage that previously used computing systems can also be easily retrofitted by a software update in order to work in the manner according to the invention. In this respect, the task is also solved by a corresponding computer program product with a computer program which can be loaded directly into a computing system, with program sections to carry out the steps of the method according to the invention, at least the steps that can be carried out by a computer, in particular the step for defining a movement model of the rail vehicle, for example by a concrete parameter selection, the step for defining sensor models, for example by a concrete parameter selection, the step for determining a Bayesian filter, the step for configuring a probability distribution of an initial system state of the rail vehicle and the step for determining a probability distribution of the pose of the rail vehicle , to be executed when the program is executed in the computing system. In addition to the computer program, such a computer program product may optionally contain additional components, such as: B. documentation and/or additional components, also include hardware components, such as hardware keys (dongles, etc.) for using the software.

A computer-readable medium, e.g. a memory stick, a hard disk or another portable or permanently installed data storage device, on which the program sections of the computer program that can be read and executed by a computer system are stored, can be used for transport to the computer system or the control unit and/or for storage on or in the computer system or the control unit. The computer system can, for example, have one or more microprocessors or the like that work together for this purpose.

The dependent claims and the following description each contain particularly advantageous refinements and developments of the invention. In particular, the claims of one claim category can also be developed analogously to the dependent claims of another claim category and their descriptive parts. In addition, within the scope of the invention, the various features of different exemplary embodiments and claims can also be combined to form new exemplary embodiments.

In one embodiment of the method according to the invention for determining the probability distribution of a pose of a rail vehicle, a Gaussian distribution is assumed for the possible deviation. Such a Gaussian distribution can be assumed in particular when small deviations are assumed, which is made possible due to the restriction of the solution space to poses that are at least roughly compatible with the track course in the track map data on which the pose determination is based. In particular, assuming small deviations and a Gaussian distribution, a particularly powerful Kalman filter (or extensions thereof) can advantageously be used, with which the computational effort is reduced and the accuracy of the estimate of the pose of the rail vehicle is particularly high.

In one embodiment of the method according to the invention for determining a pose of a rail vehicle, a linearized and decoupled motion model is used for the possible deviation, which includes an oscillatory movement of the rail vehicle around the restricted pose of the rail vehicle. Exact tracking of the probability distributions of the motion states and sensor measurement states is possible if the motion model and the sensor model are described by linear functions and the output distribution of the system is assumed to be a Gaussian distribution. Such a linear shape of the motion model and the sensor model is made possible due to the only relatively small deviations of the real pose from an inaccurate or approximate pose that can be determined by restricting the solution space.

If the individual components of the states are decoupled, it is often possible to reduce the number of possible dimensions of the deviations from the restricted possible pose in order to determine the deviations of the exact pose. The reduction in dimensions advantageously reduces the computational effort for determining the probability distributions of the motion states.

The deviations from an inclination angle of the approximate pose are preferably neglected, because the pitching tendency of a rail vehicle compared to the track geometry is usually minimal. The translational deviations in the transverse and vertical directions can also preferably be neglected, since they can be assumed to be small compared to the typically required translational accuracies. The reduction in dimensions advantageously reduces the computational effort required to determine the probability densities of the motion states.

• - ein Kalman-Filter,
• - ein Partikel-Filter,
• - ein Multi-Hypothesen-Filter.

One of the following filters is particularly preferably used to determine the pose of the rail vehicle:

• - a Kalman filter,
• - a particle filter,
• - a multi-hypothesis filter.

A Kalman filter is suitable for a linearized state system. It can typically be used for a relatively high-dimensional system, for example for a dimension n = 20. For systems that are not exactly linear, but can be appropriately linearized around the respective system state and have an approximate Gaussian distribution, the classic Kalman filter be expanded. A Kalman filter is characterized by the fact that it delivers the mathematically exact and most efficient solution under the stated conditions.

Particle filters are more suitable for solutions in systems with a low number of dimensions, as they entail a high computational effort with increasing dimension. For a system with a significantly nonlinear behavior and non-Gaussian probability distribution, the particle filter represents the most important implementation of a Bayesian filter. It is particularly flexible in use, as complex non-linear motion models and sensor models can also be processed with it. Particle filters can sometimes involve a relatively high computational effort, especially if a large number of particles are required to reproduce the distribution with sufficient accuracy. At the same time, the system dimension is significantly limited, since high-dimensional systems (for example with n > 5) can only be poorly represented by a finite number of particles.

The two most important, more complex variants of the Kalman filter are the so-called extended Kalman filter and the unscented Kalman filter. Compared to the simple Kalman filter, the extended Kalman filter and the unscented Kalman filter are more computationally intensive, since linearization around the current system state takes place at every point in time. Nevertheless, the computational effort is usually comparable to that of the simple Kalman filter, so that comparatively high-dimensional systems can also be considered here. In addition, the extensions of the Kalman filter lose the exactness of the simple Kalman filter.

A variant of the Extended Kalman filter is the Error-State Extended Kalman filter, which, in contrast to the simple Extended Kalman filter, does not track the physical system state, but instead the deviation from a nominal course of the system state. This variant has the advantage that the deviations typically only have small magnitudes and therefore the linearizations used in the extended Kalman filter have better validity, so that the pose estimation is carried out with increased accuracy.

A kind of combination of variants of the Kalman filter and the particle filter is the Rao-Blackwell particle filter, in which the system state is divided into a linear and a non-linear part, x ^l , x ^nl . In simplified terms, a particle filter can be used for the non-linear part x ^nl and a Kalman filter can be used for the linear components x ^l . The computational effort and the accuracy of the pose estimation are advantageously optimized with this hybrid approach compared to an exclusive restriction to a particle filter or a Kalman filter.

• 1 eine Darstellung von unterschiedlichen Lösungsräumen für unterschiedliche Ansätze zur Ermittlung einer Pose eines Schienenfahrzeugs,
• 2 ein Flussdiagramm, welches ein Verfahren zur Ermittlung der Wahrscheinlichkeitsverteilung einer Pose eines Schienenfahrzeugs gemäß einem Ausführungsbeispiel der Erfindung veranschaulicht,
• 3 eine schematische Darstellung einer Selbstlokalisierungseinrichtung gemäß einem Ausführungsbeispiel der Erfindung,
• 4 ein Schienenfahrzeug gemäß einem Ausführungsbeispiel der Erfindung.

The invention is explained in more detail below with reference to the attached figures using exemplary embodiments. Show it:

• 1 a representation of different solution spaces for different approaches to determining a pose of a rail vehicle,
• 2 a flowchart illustrating a method for determining the probability distribution of a pose of a rail vehicle according to an exemplary embodiment of the invention,
• 3 a schematic representation of a self-localization device according to an embodiment of the invention,
• 4 a rail vehicle according to an embodiment of the invention.

In 1 A representation 10 of different solution spaces for different approaches to determining a pose of a rail vehicle is shown.

In a diagram 10a top left in 1 A section S of a rail line is symbolically shown on which a rail vehicle 1 to be localized is located. The rail vehicle 1 can be assigned an exact pose that reflects both the current position and the current orientation of the rail vehicle 1.

At the top right of a representation 10b, the entire rectangle, which includes the section S already shown in the top left of the representation 10, is symbolized as a solution space L. That is, in this variant, the solution space L for the state of the rail vehicle 1, in particular its pose, is not restricted in advance, so that the computational effort and the inaccuracy when calculating the state of the rail vehicle 1 using a filter are very high. This variant corresponds to the procedure in J. Wendel. “Integrated navigation systems”. Munich, 2011, Oldenbourg Verlag.

Bottom left in 1 In a representation 10c, the section S already shown in the top left of the representation 10a is shown as a restricted solution space L. In this variant, the solution space L is limited to the maximum, and deviations in the pose of the rail vehicle 1 from a position and orientation of the rails cannot be taken into account. Such an approach is described in O. Heirich. "Localization of Trains and Mapping of Railway Tracks". PhD Thesis, 2020.

Bottom right in 1 10d shows a wider area around the section S as a restricted solution space L, as in the method according to the invention for determining a pose of a rail driver stuff is used. On the one hand, this restricted solution space L is sufficiently restricted, for example, to describe deviations in the orientation of a rail vehicle using a linearized model, but on the other hand, the scope is still large enough to be able to take into account typical deviations and variations in the pose of a rail vehicle in the movement model and in the sensor model.

In 2 a flow chart 200 is shown which illustrates a method for determining the probability distribution of a pose of a rail vehicle according to an embodiment of the invention.

In step 2.1, a complex model is first created, which structurally defines a system state and a movement model. Here, the system state x and the movement model BM are structured under the stipulation that a possible pose P of the rail vehicle is restricted based on the course of rail routes in track map data GKD, but a partial compensation of the restriction is based on a model-based defined possible deviation DP of the pose P of the rail vehicle from the restricted possible pose.

In step 2.II, sensor models SM are defined which describe the different measurement behavior of different sensors of the rail vehicle, which are used to determine the probability distribution p(P) of the pose P of the rail vehicle.

As a consequence of the above-described construction of the system state x, the movement model BM and the sensor models SM, a restricted solution space L results, so that the pose of the rail vehicle 1 can vary slightly relative to an orientation determined by the course of the rail tracks. This variation extends in particular to the orientation of the rail vehicle 1, since due to the degrees of freedom of the orientation of the rail vehicle 1 relative to the orientation of the rail track, a deviation of the orientation of the rail vehicle 1 from the orientation to be expected based on the map is likely. Furthermore, a deviation of the position of the rail vehicle 1 from a position estimated on the basis of a route can also occur if, for example, the car body performs lateral vibrations due to the elastic suspension.

The motion model BM can be defined with a state x as defined in equation (15) with a possible roll angle deviation and a yaw angle deviation, and a transition as shown in equation (16). The sensor model SM can be adapted as in equation (17), whereby the geocoordinates from the track map are corrected for position deviations Δy and Δz, provided that these are tracked as part of the system state.

In step 2. III, a Bayes filter BF is determined based on the system state x, the motion model BM and the sensor model SM. A Kalman filter, for example, can be used as a Bayes filter BF.

In step 2.IV, an initial system state x ₀ of the rail vehicle or its probability distribution p(x ₀ ) is now configured. For the configuration p(x ₀ ) of the system state x ₀ , for example, a previously known starting pose and its uncertainty can be used.

In step 2.V, sensor data SD are recorded by the rail vehicle’s sensors while the rail vehicle is traveling.

Furthermore, when the rail vehicle 1 is traveling in step 2.VI, updated values for the probability density p are updated based on sensor measured values z _k and the conditional probability density p(x _k |x _k-1 ) of the most recent former state x _k of the rail vehicle 1 (x _k ) of the state x _k is determined. Based on the probability density p(x) of the system state x, a current pose P of the rail vehicle 1 is continuously determined while driving.

In 3 is a schematic representation of a self-localization device 30 according to an embodiment of the invention. The self-localization device 30 has a sensor data interface 31 for receiving sensor data SD for locating a rail vehicle 1.

Furthermore, the self-localization device 30 includes a filter determination unit 32. The filter determination unit 32 is set up to determine or define a Bayesian filter BF. The Bayes filter is used as related to 2 described, based on a system state x, a movement model BM and a sensor model SM and based on track map data GKD from a geographical area in which the section of route on which the rail vehicle 1 is currently located is located. The track map data GKD is obtained from a database DB.

In the 3 In the embodiment shown, the filter determination unit 32 receives a parameter izable Bayes filter BF _p , which is based on a parameterizable system state x _p , a parameterizable motion model BM _p and a parameterizable sensor model SM _p , via interfaces (not shown in detail). The parameterization of the parameterizable Bayes filter BF _p takes place in the filter determination unit 32 by the restriction on the basis of the track map data GKD and, if necessary, on the basis of additional information I, which includes technical properties and dimensions of the rail vehicle 1, which also influence the parameterization of the Bayes filter BF _p . The additional information I can either be entered by an expert or user or obtained from an external data source or an internal data memory of the rail vehicle 1.

Part of the self-localization device 30 is also a configuration unit 33, which is set up to configure a probability distribution p(x ₀ ) of an initial system state x ₀ . The configuration unit 33 also has an interface via which an initial system state x ₀ or parameter values pw configuring the initial system state x ₀ can be entered into the configuration unit 33.

The self-localization device 30 also includes an in 3 Pose determination unit 34 marked with dashed lines. The pose determination unit 34 includes in the in 3 shown embodiment, a distribution determination unit 34a, which is set up to update values for during a journey based on sensor data SD or sensor measured values z _k and the conditional probability density p (x _k |x _k-1 ) of the latest state x _k of the rail vehicle 1 to determine the probability density p(x) (corresponds to p(x _k| z _k ) in Equation (5)) of the state x.

Furthermore, the pose determination unit 34 comprises an output interface 34b, which converts the current estimate into a format that can be used by users. In particular, an average value can be formed on the basis of a statistical distribution p(x) in order to determine a specific state x and output a pose P. For example, when implementing a particle filter, an average value can be calculated based on weighted particles, which is output as system state x or pose P.

Based on the probability density p(x) of the state x, a current pose P of the rail vehicle 1 is continuously determined while driving.

In 4 a scenario 40 is shown, which illustrates a rail vehicle 1 traveling on a rail route 2 in the direction of the arrow according to an exemplary embodiment of the invention. The rail vehicle 1 has a sensor unit 3 with which sensor data SD, for example GNSS data, which can be used to determine the status of the rail vehicle 1, are recorded. The sensor data SD is sent to a self-locating device 30, which is the in 3 has illustrated structure, transmitted. The self-locating device 30 estimates a pose P of the rail vehicle 1 based on this sensor data. This pose P is transmitted to a control unit 4. The control unit 4 controls the rail vehicle 1 based on the determined pose P.

Finally, it should be pointed out once again that the methods and devices described above are merely preferred exemplary embodiments of the invention and that the invention can be varied by a person skilled in the art without departing from the scope of the invention, insofar as it is specified by the claims. For the sake of completeness, it should also be noted that the use of the indefinite articles “a” or “an” does not exclude the fact that the characteristics in question can be present multiple times. Likewise, the term “unit” does not exclude the fact that it consists of several components, which may also be spatially distributed.

Claims (14)

Method for determining the probability distribution (p(P)) of a pose (P) of a rail vehicle (1), comprising the steps: i) defining a formal structure of a system state (x) which represents a pose (P) of the rail vehicle (1). and defining a movement model (BM) of the rail vehicle (1), which represents a probability distribution of a pose change (DP) of the rail vehicle (1) as a function of time, wherein a possible pose (P) of the rail vehicle (1) is based on the course of Rail routes (2) are restricted in track map data (GKD), but the restriction is partially compensated for based on a model-based possible deviation of the pose (P) of the rail vehicle (1) from the restricted possible pose (P), ii) defining sensor models (SM) for characterizing the measurement behavior of sensors (3) of the rail vehicle (1), which are used to determine the probability distribution (p(P)) of the pose (P) of the rail vehicle (1) based on sensor data (SD), iii) determining a Bayesian filter (BF) based on the system state (x), the motion model (BM) and the sensor model (SM), iv) configuring a probability distribution (p(x ₀ )) of an initial system state (x ₀ ) of the rail vehicle (1), v) acquiring sensor data (SD) by the sensors (3) to determine the pose (P) of the rail vehicle (1 ), vi) determining the probability distribution (p(P)) of the pose (P) of the rail vehicle (1) by applying the Bayes filter (BF) to the recorded sensor data (SD) and to the probability distribution (p(x ₀ )) the initial system state (x ₀ ) of the rail vehicle (1) as input variables. Procedure according to Claim 1 , wherein preferably during a journey of the rail vehicle (1) steps v) and vi) are repeated several times, wherein in step vi) instead of the probability distribution (p(x ₀ )) of the initial system state (x ₀ ) of the rail vehicle (1) a probability distribution (p(x _k-1 )) of a system state (x _k-1 ) last determined in connection with the last determined probability distribution (p(P)) of the pose (p(P)) of the rail vehicle (1) is used as the input variable for the Bayes filter (BF). Procedure according to Claim 1 or 2 , assuming a Gaussian distribution of possible deviation. Method according to one of the preceding claims, wherein to determine the possible deviation, a linearized and decoupled movement model (BM) for the possible deviation, which comprises an oscillatory movement of the rail vehicle (1) around the restricted possible pose (P) of the rail vehicle (1), is applied. Method according to one of the preceding claims, wherein in order to determine the possible deviation, the number of possible dimensions of the possible deviation from the restricted possible pose (P) is reduced. Procedure according to Claim 5 , where the possible deviation from an inclination angle of the restricted possible pose (P) is neglected. Procedure according to Claim 5 or 6 , where a translational deviation from the restricted possible pose (P) is neglected. Method according to one of the preceding claims, wherein one of the following filters is used as the Bayesian filter (BF): - a Kalman filter, - a particle filter, - a multi-hypothesis filter. Procedure according to Claim 8 , where the Kalman filter comprises one of the following special filter types: - an extended Kalman filter, - an unscented Kalman filter, - an error-state Kalman filter. Procedure according to Claim 8 , wherein the particle filter comprises a Rao-Blackwell particle filter. Self-localization device (30), comprising: - a filter determination unit (33) for determining a Bayesian filter (BF) by - defining a formal structure of a system state (x), which represents a pose (P) of a rail vehicle (1), and a movement model (BM) of the rail vehicle (1), which represents a pose change (DP) of the rail vehicle (1) as a function of time (t), wherein a possible pose (P) of the rail vehicle (1) is based on the course of rail routes in track map data (GKD) is restricted, but the restriction is partially compensated for based on a model-based possible deviation of the pose (P) of the rail vehicle (1) from the restricted possible pose (P), - defining sensor models (SM) to characterize the measurement behavior of sensors of the rail vehicle (1), which are used to determine the probability distribution (p(P)) of the pose (P) of the rail vehicle (1) based on sensor data (SD), - determining a Bayesian filter (BF). Basis of the system state, the movement model (BM) and the sensor model (SM), - a configuration unit (32) for configuring a probability distribution (p (x ₀ )) of an initial system state (x ₀ ) of the rail vehicle (1), - a sensor data interface ( 31) for acquiring sensor data (SD) from sensors of the rail vehicle (1) to determine the pose (P) of the rail vehicle (1), - a pose determination unit (34) for determining the probability distribution ((p(P)) of the pose (P ) of the rail vehicle (1) by applying the Bayes filter (BF) to the recorded sensor data (SD) and to the probability distribution (p(x ₀ )) of the initial system state (x ₀ ) of the rail vehicle (1) as input variables. Rail vehicle (1), comprising: - a sensor unit (3) for acquiring sensor data (SD) for determining a pose (P) of the rail vehicle (1), - a self-localization device (30) according to Claim 11 . Computer program product with a computer program which can be loaded directly into a storage unit of a control device (4) of a rail vehicle (1), with program sections to carry out the steps of a method according to one of Claims 1 until 10 to be carried out when the computer program is executed in the control device (4). Computer-readable medium on which program sections that can be executed by a computer unit are stored in order to carry out the steps of a method according to one of the Claims 1 until 10 to be executed when the program sections are executed by the computer unit.

Images