DE102021124397B4 - Method and processor circuit for determining a current vehicle pose of a motor vehicle in space and motor vehicle with the processor circuit - Google Patents

Abstract

Method for determining a current vehicle pose (19) of a motor vehicle (10), wherein the vehicle pose (19) as a pose indication (18) with a vehicle position (22) in three-dimensional space (24) and a spatial orientation ( 23) of the motor vehicle (10) at the vehicle position (22), characterized in that a driving surface (12) of the environment that extends in three-dimensional space (24) and can be driven on by the motor vehicle (10) is determined by map data as a two-dimensional multiplicity (35 ) is modeled composed of pairs of polygonal surface elements (39) connected via at least one or at most exactly one edge (40) and based on a previous pose specification (18) of the vehicle pose (19) for an incremental update of the vehicle pose (19) by a predetermined localization algorithm (25) on the basis of sensor data (26) from at least one sensor (27) of the motor vehicle (10) the respective current pose information (18) is determined by the vehicle position (22) only as a location on one of the surface elements (39) and the spatial Orientation (23) at the location on the surface element (39) is only permitted as an orientation state in a state space with a predetermined dimensionality and for this purpose in the localization algorithm (25) a solution space for the current pose indication (18) on the two-dimensional manifold (35) and the possible orientation (23) in the state space is kept limited and as a result the sensor data (26) are mapped onto the state space by the localization algorithm (25).

Description

The invention relates to a method for determining a current vehicle pose of a motor vehicle. The vehicle pose includes a location and an orientation, i.e. position information about a current location of the motor vehicle in space (vehicle position) and orientation information for specifying a spatial orientation or orientation of the motor vehicle at the location. The invention also provides a processor circuit for carrying out the method in a motor vehicle and a correspondingly equipped motor vehicle.

From the publication THOMPSON, Simon; KAGAMI, Satoshi; OKAJIMA, Masafumi. Constrained 6dof localization for autonomous vehicles. In: 2010 IEEE International Conference on Systems, Man and Cybernetics. IEEE, 2010. pp. 330-335 is a 6 spatial degrees of freedom (6DoF)) based localization method for ground-based vehicles known that captures the environment data required for the localization only by onboard sensors. The space in which the localization solution is sought is restricted by the physical limitations of ground-based vehicles, as a result of which a localization solution is calculated comparatively quickly using a GPU (Graphics Processing Unit). For the following description, two different categories of landmark-based localization algorithms are considered, 3D and 2D.

3D algorithms estimate a pose of the vehicle in 3D space. They typically estimate three degrees of position (x,y,z) and three degrees of angular freedom (pitch, roll, yaw), so a 6D state space is used. This is a very general framework that can represent any conceivable vehicle trajectory. However, it has several disadvantages. Many of the possible 6D states are unrealistic, e.g. B. could correspond to looking 60° up or flying over the ground. This makes a location algorithm less robust as it has to choose from a variety of possible vehicle poses, most of which are undesirable or implausible. There is not only a large number of unrealistic states, but also trajectories that are unrealistic overall. An example is a multi-story parking garage where a vehicle travels on one floor for a period of time without encountering any landmarks. This blind estimation of the vehicle's pose could result in a drift in altitude, causing the vehicle to suddenly appear to be on a different floor of the building. While it is clearly impossible for the vehicle to drive through the ceiling, this limitation is difficult to encode in a 3D frame.

Another problem is the visibility of landmarks or landmarks. Many landmarks are not visible from a given position due to occlusion. It is often not possible to perform such occlusion calculations online in real time, but they could be performed offline beforehand. However, this is very difficult when the state space is all 3D space, so a list of visible landmarks has to be computed for every possible 3D position.

2D algorithms estimate a pose on a 2D surface, so the state space consists of two position coordinates and the yaw angle and is thus three-dimensional. This is a good representation for most common driving situations with a motor vehicle, as it usually moves along a ground plane. However, there are situations where a global ground plane is not a good approximation, e.g.: multi-storey car parks with ramps connecting the floors, or highway bridges crossing other major thoroughfares with ramps leading up and down the highway roads at a steep angle .

A method for locating a motor vehicle on the basis of odometry data from at least one sensor of a motor vehicle is known from WO 2020/248209 A1. It describes how the three-dimensional space surrounding the motor vehicle can be described by map data as a two-dimensional horizontal surface (x, y), in which each surface point is assigned four additional coordinates or dimensions, namely height information (z) and three orientation angles for Yaw, pitch and roll of the motor vehicle. The disadvantage of this method is that multi-storey buildings, such as multi-storey car parks, are difficult to model because there are horizontal surfaces on top of each other. In addition, the height information (z) must be stored as dimension information for each surface point, which is why the map data requires correspondingly large amounts of data.

From the U.S. 8,531,456 B2 It is known to simulate a surface extending in three-dimensional space by means of a multiplicity of polygonal surface elements, for example triangles, adjoining one another, with which the surface can be represented in three-dimensional space as a two-dimensional multiplicity. The arrangement of the polygonal surface elements is also referred to as a mesh. For example, a Sta do be modeled as a three-dimensional model of the statue surface.

The object of the invention is to provide an efficient method for locating a motor vehicle in three-dimensional space, by means of which it is possible to specify the vehicle position and the spatial orientation of the motor vehicle, ie the pose specification of a vehicle pose.

The object is solved by the subject matter of the independent patent claims. Advantageous developments of the invention are described by the dependent patent claims, the following description and the figures.

As a solution, the invention includes a method for determining a current vehicle pose of a motor vehicle, the vehicle pose being determined by a processor circuit as a pose specification with a vehicle position (vehicle location) in three-dimensional space and a spatial orientation of the motor vehicle (vehicle orientation) at the vehicle position. The pose specification can therefore be four-dimensional, for example, if the vehicle positions by spatial coordinates x, y, z and the orientation provides a yaw angle or an alignment (English: heading) to describe the spatial orientation of the motor vehicle. Further possible orientation information can be the pitch angle and/or the roll angle. In the further course of the description, further possible information for orientation is described.

In order to cyclically update the pose information during a journey of the motor vehicle, the method includes using map data to map a driving surface (driving surface) of the environment or the surroundings, which extends in three-dimensional space and can be driven on by the motor vehicle, as a two-dimensional multiplicity of pairs over at least one or at most exactly one edge of connected polygonal surface elements is modeled as a composite. In other words, the driving surface, i.e. the driving surface or the roadway, is described or modeled by the polygonal surface elements described at the beginning, i.e. the driving surface is modeled as a mesh. In a multi-storey car park, each floor can thus be modeled by flat or horizontally arranged surface elements, and ramps for changing between floors can be modeled as inclined surface elements or arranged obliquely in space. Unevenness, such as speed bumps, can also be modeled as a three-dimensional curved area of the driving surface.

In each cycle, the cyclic updating of the pose specification can build on the result of the previous cycle, ie the previous pose specification, or develop it further. For example, the odometry data described at the outset can be received, which indicate by what distance and with what orientation (heading) the motor vehicle has moved since the last pose indication. The method accordingly includes that starting from a previous pose indication of the vehicle pose for an incremental update of the pose indication by a predetermined localization algorithm on the basis of sensor data of at least one sensor of the motor vehicle the respective current pose indication is determined. A localization algorithm as known from the prior art can be used here. Starting from the previous pose specification, for example, a vector or a line can be constructed or determined, which extends or runs in space according to the determined heading. In this case, however, impossible or implausible results must be avoided, such as those that can be caused by measurement scatter in the sensor. For example, it must be noted that there cannot be a wall between the previous pose specification and the new pose specification, which the motor vehicle would have to have driven through in order to arrive at the newly determined vehicle pose. Accordingly, the localization algorithm must be limited or operated with boundary conditions that avoid implausible estimation results for the estimation of the current vehicle pose. This is implemented in the method according to the invention, in which the vehicle position is only permitted as a location on one of the surface elements and the spatial orientation at the location on the surface element is only permitted as an orientation state in a state space with a predetermined dimensionality, and for this purpose in the localization algorithm a solution space for the current pose information is kept limited to the two-dimensional multiplicity and the possible orientation in the state space, and as a result the sensor data are mapped onto the state space by the localization algorithm. The "state space" used herein represents the complete description of the state of the motor vehicle in relation to its arrangement in space, ie the state space includes both the indication of the location of the motor vehicle on the two-dimensional manifold and the spatial alignment/orientation at the location. The state space is thus at least two-dimensional in order to describe the position of the motor vehicle on the manifold (without further information on the spatial alignment at the location). However, in addition to the location, the state of orientation is also present specified, so that the state space has a dimensionality (number of dimensions) greater than 2 in order (as will be described below) to specify, for example, an additional orientation state such as a yaw angle or orientation angle (heading). The at least one additional dimension that goes beyond the location on the manifold and describes the state of orientation is also referred to as a mathematical fiber.

In other words, the plausibility check includes two measures. On the one hand, only a movement of the motor vehicle within the area of the manifold, ie within the adjoining surface elements, is permitted. If, for example, the odometry data signals that the motor vehicle has moved a certain distance, e.g. five meters, and this shows that the motor vehicle must have left a horizontal surface element via an edge of the same, which is adjoined by a surface element leading diagonally upwards, the route covered with the signaled route length, i.e. five meters, for example, is mapped onto the horizontal surface element in the first section, then bent at the edge according to the angle of inclination of the adjacent surface element and the remaining part of the route length is continued there in the inclined surface element until the End of the route length, for example five meters, is reached. This results in two computationally relatively simple operations on two adjacent surface elements. With regard to the spatial orientation of the motor vehicle, the inclined position of the respective surface element already signals which orientation the motor vehicle must have based solely on its position on the surface element, ie the inclination or the gradient of the surface element. This additional information results from the use of surface elements that can have an orientation or position in three-dimensional space. As a second measure, the spatial orientation, e.g. based on the yaw rate, is measured and thus it can also be determined from the spatial position of each surface element how the motor vehicle is absolutely aligned in space.

The advantage resulting from the invention is that the localization can be carried out on the basis of geometric calculations which are limited to surface elements and the defined state space of the orientation, which advantageously reduces the computational complexity. In addition, the restriction to the surface elements and the predefined state space for the spatial orientation can ensure that a plausible localization result always results.

The invention also includes developments that result in additional advantages.

A development includes that when the location on one of the surface elements is determined by the localization algorithm, a transition between two surface elements is only permitted via a respective common edge of the surface elements. In other words, a change between two surface elements is only possible if they are connected via a common edge. This can prevent the localization algorithm from determining a current pose specification that would in reality require driving through a closed wall. In other words, it is possible to prevent implausible position information from being determined in the modeling with little effort by modeling impossible or non-existent routes by using surface elements at a distance from one another, i.e. without a common edge, describing them or entering them in the map data.

A development includes that the trafficable driving surface has at least one convex and/or at least one concave curvature and the driving surface is modeled by defining the surface elements as flat polygons. In other words, each surface element is flat or not curved, ie the corners or corner points of the respective polygon of a surface element are arranged in a coplanar manner. A concave or convex curvature in the running surface itself is approximated by several flat surface elements, e.g. by arranging them at an acute or obtuse angle to one another (angles other than 180°). The accuracy of this approximation can be adjusted as required by the size of the surface elements. For example, an outer dimension of a surface element can be selected depending on the curvature of the roadway surface. A surface element can, for example, represent a surface section of the driving surface with an external dimension in a range from 30 centimeters to 10 meters, to name suitable examples. By also approximating or modeling a concave or convex curvature by multiple planar polygons, the localization algorithm is prevented from reaching a computational complexity greater than an allowable maximum. For example, the size of the surface elements can be specified as a function of the permissible computational complexity.

A development includes that a distance calculation between two specified Points on the two-dimensional manifold is computed as a line length of a line located solely within the surface elements. In particular, when calculating a route length on or along a convex or concave curvature of the driving surface, this prevents only the straight line, i.e. the Euclidean distance between a starting point and an end point, from being calculated, but rather the course of the curvature is determined by the course of the route replicated or approximated within the surface elements. This gives a more accurate calculation or indication of the route length.

A further development includes at least some of the sensor data including odometry data in order to determine the pose information in a roofed building in the absence of a position signal from a GNSS (global navigation satellite system). In the manner described, the odometry data can specify a distance covered since the last measurement/the last cycle, i.e. since the last preceding pose specification, and a direction of travel or an orientation (heading), with the direction of travel/the heading being recorded over time as a temporal progression of the Heading can be specified for different points in time. For example, the heading can be specified as steering angle information and/or cardinal direction. Corresponding sensors in the chassis and/or an electronic compass can be provided as a sensor in the vehicle. Thus, for example, in a parking garage where the position signal of a GNSS, for example the GPS (global positioning system) is blocked or shielded due to the existing building ceilings and can therefore be missing, the vehicle position in three-dimensional space (including height information) can be determined using the algorithm.

A development includes that for the respective surface element, texture data on a texture of the environment are provided in the surface element. By dividing the runway into surface elements of a predetermined surface size, which is in particular larger than 0.5 square meters, the storage effort for specifying the nature of the environment on the respective surface element is also limited. For example, a surface element can cover an area in a range from 0.5 square meters to 30 square meters. Then only quality data for each of these surface units is necessary in order to store the quality of the environment on this surface element in the map data for the entire surface element.

A development includes that the texture data indicate which surrounding objects or landmarks are potentially detectable from the respective surface element. In the localization algorithm, an object detection and/or an object classification and/or a matching algorithm is limited to the respectively potentially detectable landmarks. A landmark can be a stationary object with a known appearance (related to the respective measured physical quantity, e.g. light, radar, ultrasound) and known position (e.g. geoposition and/or 3D position in space and/or orientation). If it is determined on which surface element the motor vehicle is located, the property data can indicate which landmarks are potentially "visible" from there if they are not currently covered, for example by another road user or by a weather phenomenon such as fog or heavy rain. In the localization algorithm, for example, a machine learning model, for example an artificial neural network, can then be used for object detection (checking whether a specific object is present) and/or object classification (determining a class of the object, for example pedestrians or traffic signs) for those Be limited to landmarks or object classes that are only potentially or theoretically visible or detectable from the surface element.

This increases the robustness of object detection and object classification. Additionally or alternatively, the assignment of the object detections to landmarks from the map (i.e. the matching) can be restricted.

The current vehicle pose can be determined in two stages from previous vehicle poses. A further development for this includes that the localization algorithm from the at least one sensor receives both odometry data of a route covered since or since the previous pose indication and direction of travel as well as environmental data of surrounding objects sensed or recognized at the current vehicle position and initially on the basis of the route and direction of travel intermediate position information is determined and, based on this, the current position information is determined by comparing the environmental data with the detectable landmarks. The localization is therefore carried out in two stages, in that first an intermediate position or rough or provisional position information of the vehicle position (intermediate current position information) is determined using the odometry data and then, for example, on the basis of camera data or general sensor data that signal the area around the motor vehicle (environment data). the object recognition and/or object classification already described, the relative position of the environment projects is determined. Since position data corresponding to these surrounding objects can be stored in the map data, ie landmarks are described, the current vehicle pose can be inferred or calculated by means of triangulation, for example.

A development includes that the condition data indicate a ground condition of the driving surface and the ground condition is signaled to a driver assistance function of the motor vehicle. For example, a coefficient of friction and/or an indication of a material of the ground (concrete or gravel or bitumen) can be signaled to the driver assistance function as the ground condition. Based on this, a driver assistance function can, for example, adapt a corresponding configuration for an anti-lock braking function in order to take the condition of the ground into account.

A further development includes that the location on the surface element as an ID of the surface element and a coordinate in a coordinate system of the surface element are specified as internal pose information, and the location is converted into absolute coordinates of the three-dimensional space for a control unit receiving the current position information according to the internal pose information and a travel trajectory for automated travel on the driving surface is planned by the control unit using an automated and/or assisted driving function, and the motor vehicle is accordingly guided on the driving surface by means of driving actuators. A compact representation of the pose specification can thus be specified internally in the processor circuit when the method is carried out by ID (identification specification of the respective surface element) and the coordinates within the surface element. On the other hand, for every external control unit that needs the current pose information or is supplied with it, the usual signaling using absolute coordinates in three-dimensional space is possible, so that the control unit does not have to be adapted to the modeling based on the surface elements, but the control unit with it without further adaptation the method and the processor circuit can be combined.

• • eine Dimension für eine Gierwinkelangabe oder
• • zwei Dimensionen für eine Gierwinkelangabe sowie eine zusätzliche Wankwinkelangabe oder Nickwinkelangabe oder
• • drei Dimensionen für drei Raumwinkelangaben

umfasst. Die drei Raumwinkelangaben können Gierwinkel, Wankwinkel und Nickwinkel angeben. Eine Kombination aus einer Position in einer 2D-Mannigfaltigkeit und einem an dem jeweiligen Punkt oder der jeweiligen Position definierten zusätzlichen Zustandsraum ist in der Mathematik auch beschrieben als Mannigfaltigkeit mit Fasern, wobei der zusätzliche Zustandsraum die Fasern darstellt. Durch die Beschränkung auf die Dimensionen ergibt sich der Vorteil, das auch die Orientierung robuster geschätzt werden kann. Der Zustandsraum kann zusätzlich zu den hier angeführten Winkeln oder Winkelangaben noch die Position auf der Mannigfaltigkeit (ID des Flächenelements und 2D-Position/Koordinaten auf dem Flächenelement) als weitere Dimensionen umfassen. Der Zustandsraum ist also für die angeführten Beispiele drei, vier oder fünfdimensional, d.h. die Dimensionen der Mannigfaltigkeit (2) und die Dimensionen der Fasern (1, 2 oder 3).A further development includes that the state space with the predetermined dimensionality:

• • a dimension for a yaw angle specification or
• • two dimensions for a yaw angle specification and an additional roll angle specification or pitch angle specification or
• • three dimensions for three spatial angle specifications

includes. The three solid angle specifications can indicate yaw angle, roll angle and pitch angle. A combination of a position in a 2D manifold and an additional state space defined at the respective point or position is also described in mathematics as a manifold with fibers, where the additional state space represents the fibers. The restriction to the dimensions results in the advantage that the orientation can also be estimated more robustly. In addition to the angles or angle specifications listed here, the state space can also include the position on the manifold (ID of the surface element and 2D position/coordinates on the surface element) as further dimensions. The state space is therefore three, four or five-dimensional for the given examples, ie the dimensions of the manifold (2) and the dimensions of the fibers (1, 2 or 3).

• • eine Dimension für eine Gierwinkelangabe oder
• • zwei Dimensionen für eine Gierwinkelangabe sowie eine zusätzliche Wankwinkelangabe oder Nickwinkelangabe oder
• • drei Dimensionen für drei Raumwinkelangaben

darstellen oder belegen. Insbesondere ist die Orientierungsangabe auf jeweils genau diese Dimensionen beschränkt. Die drei Raumwinkelangaben können Gierwinkel, Wankwinkel und Nickwinkel angeben. Die Dimension des Zustandsraums ergibt sich als die Summe der Dimensionen der Mannigfaltigkeit selbst (als Basismannigfaltigkeit für die Ortsangabe (zwei für eine zweidimensionale Mannigfaltigkeit)) und der Dimensionen der Faser (eins, zwei oder drei) und ist also für die angeführten Beispiele drei, vier oder fünf. Durch die Beschränkung der Dimensionen ergibt sich der Vorteil, dass auch die räumliche Orientierung robuster geschätzt werden kann. Das Reduzieren von Sensordaten auf die nur verfügbaren Dimensionen des Faserbündels oder der Fasern kann z.B. durch eine mathematische Projektion erfolgen.A development includes that the state space consists of a combination of a position in the 2D manifold and additional data for the orientation of the vehicle. In mathematics, such a configuration is referred to as a "fiber bundle over a manifold". The additional data, also called fibers, can be in spaces with the predetermined dimensionality:

• • a dimension for a yaw angle specification or
• • two dimensions for a yaw angle specification and an additional roll angle specification or pitch angle specification or
• • three dimensions for three spatial angle specifications

represent or prove. In particular, the indication of orientation is limited to precisely these dimensions in each case. The three solid angle specifications can indicate yaw angle, roll angle and pitch angle. The dimension of the state space results as the sum of the dimensions of the manifold itself (as the base manifold for the location specification (two for a two-dimensional manifold)) and the dimensions of the fiber (one, two or three) and is therefore three, four for the given examples or five. The limitation of the dimensions results in the advantage that the spatial orientation can also be estimated more robustly. The reduction of sensor data to the only available dimensions of the fiber bundle or the fibers can be done, for example, by a mathematical projection.

A development includes that the status space includes status information of a chassis of the motor vehicle, which is signaled by at least one chassis sensor. For example, due to different positions of the shock absorbers and/or the suspension of the individual tires of the motor vehicle, such as when driving over an uneven surface and/or an object such as a sidewalk or a speed bump. This state information can also be combined with one of the state spaces described above, whereby the corresponding number of dimensions (dimensionality) of the state space is expanded.

As a further solution, the invention includes a processor circuit for a motor vehicle, the processor circuit being set up to carry out an embodiment of the method. The processor circuit can be provided, for example, by a control unit of the motor vehicle. The processor circuit can include at least one microprocessor and/or at least one microcontroller and/or at least one digital signal processor (DSP) and/or at least one ASIC (application specific integrate circuit). The processor circuit can also include the data memory, in which computer-readable program instructions can be stored which, when executed by the processor circuit, cause the processor circuit to carry out the embodiment of the method. For this purpose, it can be coupled to the respective microprocessor/microcontroller/DSP/ASIC.

As a further solution, the invention comprises a motor vehicle with at least one sensor and with an embodiment of the processor circuit which is coupled to the at least one sensor. The motor vehicle can be designed as a passenger car or truck, for example. A camera, a radar, a LIDAR, an ultrasonic sensor can be provided as a sensor.

The invention also includes the combinations of features of the described embodiments.

• 1 eine schematische Darstellung einer Ausführungsform des erfindungsgemäßen Kraftfahrzeugs;
• 2 eine Skizze zur Veranschaulichung von Kartendaten, in welchen eine befahrbare Fahrfläche als zweidimensionale Mannigfaltigkeit im dreidimensionalen Raum durch ebene Flächenelemente modelliert ist.
• 3 eine Skizze zur Veranschaulichung einer Berechnung einer Streckenlänge bei einer konkaven Krümmung der Fahrfläche;
• 4 eine Skizze zur Veranschaulichung einer Ausführungsform des erfindungsgemäßen Verfahrens.

Exemplary embodiments of the invention are described below. For this shows:

• 1 a schematic representation of an embodiment of the motor vehicle according to the invention;
• 2 a sketch to illustrate map data in which a driving surface is modeled as a two-dimensional manifold in three-dimensional space by flat surface elements.
• 3 a sketch to illustrate a calculation of a route length with a concave curvature of the driving surface;
• 4 a sketch to illustrate an embodiment of the method according to the invention.

The exemplary embodiments explained below are preferred exemplary embodiments of the invention. In the exemplary embodiments, the described components each represent individual features of the invention to be considered independently of one another, which also develop the invention independently of one another and are therefore also to be regarded as part of the invention individually or in a combination other than that shown. Furthermore, the exemplary embodiments described can also be supplemented by further features of the invention already described.

Elements with the same function are each provided with the same reference symbols in the figures.

1 shows a motor vehicle 10, which can be, for example, a passenger car or truck. The motor vehicle 10 can perform a journey 11 over a driving surface or a driving surface 12 . For example, motor vehicle 10 may be traveling across a street or across the floor of a parking garage. In this case, a control unit 13 can operate an automated and/or assisted driving function 14 (e.g. a partially autonomous or fully autonomous driving function and/or a traffic light assistant), which can carry out longitudinal guidance and/or lateral guidance of motor vehicle 10 while driving 11 by Control signals are generated for corresponding driving actuators 15, for example for a steering and/or brake and/or a drive motor, in order to guide motor vehicle 10 along a trajectory 16 calculated by control unit 13. To this end, control unit 13 can access map data 17 from a navigation database DB. Furthermore, current pose information 18 of a current vehicle pose 19 can be signaled to control unit 13 by a processor circuit 20 of another control unit 21 or by a processor circuit that can be integrated in control unit 13 (not shown). The vehicle pose 19 can specify a current vehicle position 22 and a spatial orientation 23 of the motor vehicle 10 in the three-dimensional space 24 (eg as an angle specification with respect to a predetermined reference angle). The processor circuit 20 can execute a localization algorithm 25 for cyclically determining the respective current pose indication 18 , which cyclically receives sensor data 26 from at least one sensor 27 can. A wheel odometry or vehicle odometry 28 can provide corresponding odometry data 29 as sensor data 26 as a possible sensor 27 . As a possible sensor 27 , a camera 30 can provide surroundings data 31 of surroundings 32 of motor vehicle 10 as sensor data 26 . For this purpose, a detection range 33 of the camera 30 can be aligned with the surroundings or the environment 32 , for example in a front region 34 located in front of the motor vehicle 10 . In addition or as an alternative to the camera 30, at least one further sensor for generating environmental data 31 can be provided. In addition or as an alternative to vehicle odometry 28 , a receiver of a position signal from a GNSS can be provided, for example, in order to provide sensor data 26 that can be used in addition or as an alternative to odometry data 29 .

The localization algorithm 25 ensures that the current pose specification 18 signals a plausible or realistic vehicle pose 19, for example by preventing a vehicle position 22 from being specified that would require the motor vehicle 10 to hover in front of the driving surface 12 or in the Runway 12 is sunk.

The localization algorithm 25 can provide a cyclical and/or iterative update of the current pose indication 18 in that the current sensor data 26 is received in successive measurement cycles or localization cycles and the current pose indication 18 is estimated based on a previously estimated vehicle pose.

For this purpose it can be provided that the map data 17 from the navigation database DB models the driving surface 12 in the three-dimensional space 24 as a two-dimensional manifold.

2 illustrates the information from such map data 17, in which a two-dimensional manifold 35 can model the driving surface 12 in a parking garage 36, i.e. the different floors 37 and, for example, ramps 38 for changing between the floors 37. The manifold 35 can be modeled from individual surface elements 39 be, which are formed here as plane triangles. Instead of a triangle, another polygon, for example squares or hexagons, can also be used as a basis. The surface elements 39 are in particular flat, that is to say surfaces which are free of curvature. A closed driving surface 12 can be formed by connecting two surface elements 39 via a common edge 40 of the polygons, which can signal to the localization algorithm 25 that a continuous driving movement of the trip 11 on the driving surface 12 is possible via this edge 40. Corresponding map data 17 can be formed, for example, on the basis of the algorithm described above.

3 shows in a sectional view of the real driving surface 12 how a concave curvature 41 of the driving surface 12 can be simulated by individual surface elements 39 and their common edges 40. For the localization algorithm 25, this means that a route length between a first point 42 (starting point) and a second point 43 (end point) does not have to be calculated as a Euclidean distance 44 (Euclidean distance), but rather the route length from straight route sections 45 can be calculated, one of which is located in one of the surface elements 39. If, for example, the odometry data 29 signals that the motor vehicle has covered a distance of a given length since the previous pose indication, then the Euclidean distance 44 does not have to be assumed, but the length of the route from the route sections 45 and their length within the surface elements 39 can be calculated more precisely be taken as a basis.

4 illustrates how the localization algorithm in a step S10 can receive sensor data 26, on the one hand, odometry data 29 of a vehicle movement during the journey 11, based on a previous pose indication 46, a taxi route or route 47 and heading information 48, and as environmental data 31, for example, a camera image 49. In a step S11, based on the previous pose information 46, the route 47 can be mapped within the surface elements 39 according to the heading information 48, so that the odometry data 29 of the sensor data 26 are mapped onto the two-dimensional multiplicity 35 in space 24. This can result in a provisional or intermediate pose indication 50 . Proceeding from this, in a step S12 on the basis of the environment data 31, the final current

Pose indication 18 can be determined. For this purpose, for example, it can be determined from the map data for the current surface element 51, which can be identified by the intermediate pose specification 50, which texture data 52 results from the map data 17. The texture data 52 can, for example, identify those landmarks 53 that are potentially or at all only visible from the current surface element 51 or can be censored or detected by the sensors 27, ie are in a line of sight with respect to the camera 30, for example. Based on the environment data 31 and the condition data 52 relating to the landmarks 53, the current pose indication 18 can then be determined, for example, on the basis of triangulation. A corresponding comparison of landmarks 53 in environmental data 31 can be implemented or realized using a method from the prior art, for example using a method of machine learning for object detection. The current pose information 18 can include, for example, a respective identity information ID of the respective surface element 39 (in the example the surface element F1) and local coordinates U, V of the determined vehicle position 22 within the surface element 51 (here within F1).

This ID and the coordinates U, V can then be converted into absolute coordinates in space 24 in order to describe or signal the vehicle position 22 in the current pose specification 18 for the control device 13 . Thus, the use of the manifold 35 by the localization algorithm 25 for the control unit 13, in particular the automated driving function 14, is transparent or not visible, so that the control unit 13 and/or the driving function 14 for interpreting the current pose information 18 does not have to have been adapted or changed.

The position of a vehicle can thus be determined on the basis of a map of the surroundings and an image stream from a camera system, possibly with additional sensor data such as GNSS (Global Navigation Satellite System, e.g. GPS), INS (Inertial Sensor) or vehicle odometry (e.g. wheel revolutions, steering angle , compass data) can be determined. This is a core task for autonomous driving, but also for various driver assistance systems, i.e. for automated and/or assisted driving in general. The localization can be carried out by comparing landmarks stored in the map (e.g. traffic signs, lane markings) with objects extracted from the camera image using a machine learning algorithm, and the vehicle pose can be calculated from this.

The idea proposes an approach that combines the robustness of 2D approaches with the generality of 3D approaches. The basic idea is to define the state space on a two-dimensional surface (a manifold) embedded in the three-dimensional space. According to the localization algorithm, the vehicle is forced to move on this manifold and can therefore only reach physically possible positions and state transitions.

A definition of a manifold is limited here to the present case. A 2D manifold is therefore a subset of 3D space that locally looks like a planar surface element. The important keyword is "local", meaning that while there is always a local tangent plane to move on, a manifold can globally encode all navigable areas, including the ramps of a multi-story garage and/or freeway overpasses. A surface element can be represented as a triangular mesh, i.e. as a set of triangles embedded in 3D space, where each triangle must be adjacent to at most one other triangle along each of its edges.

• • eine S1-Faser, bei der wir den Gierwinkel als zusätzliche Information haben,
• • eine S2-Faser, bei der wir einen Gier- und einen Neigungswinkel als zusätzliche Information haben. Dies würde die Kabinenposition auf den Basis-Mannigfaltigkeit beschränken, aber eine Neigung der Kamera ungleich Null erlauben, z. B. beim Überqueren einer Bodenschwelle oder Bremsschwelle,
• • eine SO(3)-Faser, bei der wir einen vollständigen Gier-, Nick- und Rollwinkel der Kamera haben.

The state space of a vehicle is a so-called fiber bundle over the base manifold. That is, the pose of the vehicle is described by a point on the manifold (the triangle of the mesh it belongs to and the coordinates on the triangle) together with an additional datum encoding the orientation. Possible fibers are:

• • an S1 fiber where we have the yaw angle as additional information,
• • an S2 fiber where we have a yaw and a pitch angle as additional information. This would constrain the cabin position to the base manifold, but allow for non-zero tilt of the camera, e.g. B. when crossing a speed bump or speed bump,
• • an SO(3) fiber where we have full camera yaw, pitch and roll angles.

It is also possible to extend the state space further by abandoning the assumption that the vehicle is a rigid body. In this case, one could add internal states of the vehicle, such as the expansions of the shock absorbers, to the fibers. For the fiber bundle S1, every possible state of the vehicle is physically realistic since the vehicle is on the base manifold and oriented parallel to it, so the first problem of the 3D algorithms does not apply here. For the larger fibers one would add some constraints that avoid unreasonable pitch and roll angles. In the case of a dynamic vehicle model, it would even be possible to model complex relationships between shock absorber states and/or camera orientations and/or the local structure of the manifold.

Transitions to update location/pose indication: Preferably, the location algorithm begins with an initial pose estimate, followed by incremental updates as the vehicle moves. In the framework described, only incremental updates are preferably allowed, which end on the same or a neighboring surface element (eg triangle), and thus the connectivity of the network (mesh) of the surface elements encodes possible state transitions. In this way, physically impossible transitions such. B. driving through a wall or through the ceiling, is not possible due to the construction and thus the second problem of the 3D algorithms does not occur.

Collateral Information: Since the vehicle is always on a specific triangle/area of the mesh, the triangles can be used to store additional information that facilitates localization. The connectivity and hence the possible movements are already implicitly stored in the manifold structure, but we can also store for each triangle/area element: a list of landmarks potentially visible from that triangle. This can be calculated offline when the map is generated and one can avoid expensive occlusion calculations. With this list, obscured or non-existent landmarks are avoided from being presented to the matching algorithm, making the matching algorithm both faster and more robust. The type of surface the vehicle is currently on (ramps, parking lots, taxi lanes) as well as regulatory data (speed limits, one-way/two-lane restrictions, maximum clearances). This information can be used to adapt the vehicle's behavior during automated and/or assisted driving, or it can be displayed to the driver in the context of a driver assistance system.

It is theoretically also possible in the prior art to store such information for a 3D algorithm as well, but there you would need an additional finite partitioning of the space and search structures in order to access the data, while here these structures are obtained free of charge from the network.

Use sensor data for localization: The described manifold framework also allows an improved integration of 2D sensor data. An example is the speed information of wheel odometry. This can be used as a boundary condition for the 3D distance between two adjacent sensor positions. However, when the vehicle moves on a curved surface, this leads to a systematically overlong result, as in 3 shown. If the information on the manifold is available, velocity can be used instead as a measure of the geodetic distance on or in the manifold, yielding a model that better reflects the situation. 3 illustrates motion on a curved surface (side view). A 3D algorithm would take the wheel odometry speed reading as a measure of the length of the straight line (minimum distance 44 between the two points in 3D), when correctly it would have to be the geodetic distance on the driving surface.

The localization algorithm: Each state on our fiber bundle preferably corresponds to a full 6D state in the surrounding 3D space. Therefore, most of the pipeline can be reused from well-known 3D localization algorithms, such as landmark “matching” to 2D detections (however, it can optionally be enhanced by using the triangle-based side information as described above). There are two places where a given 3D algorithm must be modified: Given an initial pose for the 3D localization algorithm, a suitable nearby pose on the manifold must be found by projection onto the manifold. Alternatively, localization algorithms can be developed that calculate the initial pose directly on the manifold. The vehicle pose update, which usually solves an optimization problem where the vehicle pose is the variable, must be modified to take into account the structure of the manifold. In particular, the optimization must be able to handle transitions from one triangle to another. With these modifications, any existing landmark-based 3D localization algorithm can be modified to work on a manifold with little effort.

Interaction with a propulsion controller (controller 13): While the representation on the manifold allows for more robust localization, it also has additional applications for a propulsion controller. The local structure of the world or environment (e.g. the curvature of the road) is available and one can also potentially estimate internal states of the vehicle such as shock absorber states. This information can be passed on to a drive controller for autonomous driving, for example, so that it can be used to adjust the steering angle and acceleration.

An extension of the landmark-based 3D localization algorithms results, which restricts the state space to a fiber bundle over a two-dimensional manifold defined by a triangular mesh or generally planar surfaces chen elements is shown. While this is flexible enough for any possible trajectory of a vehicle, it offers the following advantages over a full 3D algorithm. It constrains the state space and therefore forces the localizer to produce physically meaningful indications of vehicle poses. This makes the algorithm more robust. By using the connectivity information of the manifold (edge transitions), impossible trajectories e.g. B. go through walls or ceilings, excluded. This increases the stability of the algorithm, especially in long dead reckoning phases without visible landmarks. The mesh can be used to store additional information with each triangle, such as a list of potentially visible landmarks. This increases both the robustness and the speed of the matching process.

The geometry of the manifold allows better use of sensor readings such as wheel odometry, which measures geodetic distance instead of 3D distance, and can also serve as an additional input for a propulsion controller that can perform steering movements as a function of local curvature.

Overall, the examples show how a vehicle localization can be provided using a state space for a 2D manifold and further dimensions for the indication of orientation.

Reference List

10

motor vehicle

11

Drive

12

driving surface

13

control unit

14

driving function

15

driving actuators

16

trajectory

17

card data

18

pose statement

19

vehicle pose

20

processor circuit

21

control unit

22

vehicle position

23

orientation

24

three-dimensional space

25

localization algorithm

26

sensor data

27

sensor

28

odometry

29

odometry data

30

camera

31

environment data

32

Environment

33

detection range

34

front area

35

manifold

36

parking garage

37

Floors

38

ramps

39

surface elements

40

edge

41

curvature

42

Point

43

Point

44

distance

45

track sections

46

pose statement

47

driving distance

48

heading information

49

camera image

50

pose statement

51

surface element

52

texture data

53

landmarks

Claims (14)

Method for determining a current vehicle pose (19) of a motor vehicle (10), wherein the vehicle pose (19) as a pose indication (18) with a vehicle position (22) in three-dimensional space (24) and a spatial orientation ( 23) of the motor vehicle (10) at the vehicle position (22), characterized in that a driving surface (12) of the environment that extends in three-dimensional space (24) and can be driven on by the motor vehicle (10) is determined by map data as a two-dimensional multiplicity ( 35) composed of polygonal surface elements (39) connected in pairs via at least one or at most precisely one edge (40). is modeled and based on a previous pose specification (18) the vehicle pose (19) for an incremental update of the vehicle pose (19) by a predetermined localization algorithm (25) on the basis of sensor data (26) of at least one sensor (27) of the motor vehicle (10 ) the respective current pose information (18) is determined by the vehicle position (22) only as a location on one of the surface elements (39) and the spatial orientation (23) at the location on the surface element (39) only as an orientation state in a state space with a predetermined dimensionality is permitted and for this purpose in the localization algorithm (25) a solution space for the current pose indication (18) is kept limited to the two-dimensional multiplicity (35) and the possible orientation (23) in the state space and as a result the sensor data (26) is Localization algorithm (25) are mapped to the state space. procedure after claim 1 , wherein when the location on one of the surface elements (39) is determined by the localization algorithm (25), a transition between two surface elements (39) is permitted only via a respective common edge (40) of the surface elements (39). Method according to one of the preceding claims, wherein the trafficable driving surface (12) has at least one convex and/or at least one concave curvature (41) and the driving surface (12) is modeled by defining the surface elements (39) as planar polygons. Method according to one of the preceding claims, in which a distance calculation between two predetermined points (42, 43) on the two-dimensional manifold (35) is calculated as a line length of a line located exclusively within the surface elements (39). Method according to one of the preceding claims, wherein for determining the pose indication (18) in a roofed building in the absence of a position signal from a Global Navigation Satellite System, GNSS, at least part of the sensor data (26) is odometry data (29). Method according to one of the preceding claims, wherein texture data (52) of a texture of the environment in the surface element (39) are provided for the respective surface element (39). procedure after claim 6 , wherein the texture data (52) indicates which landmarks (53) can potentially be detected from the respective surface element (39) and in the localization algorithm (25) an object detection and/or an object classification and/or a matching algorithm is applied to the respective potentially detectable landmarks (53) is limited. procedure after claim 7 , whereby the localization algorithm from the at least one sensor (27) generates both odometry data (29) of a route (47) and direction of travel traveled since or since the previous pose indication (18) and environmental data (31) from the current vehicle position (22) sensed objects in the surroundings are received and initially based on the route (47) and direction of travel an intermediate current position information and based on this by a comparison of the environment data (31) with the detectable landmarks (53) the current pose information (18) is determined. Procedure according to one of Claims 6 until 8th , wherein the condition data (52) indicate a ground condition of the driving surface (12) and the ground condition is signaled to a driver assistance function of the motor vehicle (10). Method according to one of the preceding claims, wherein an internal pose specification specifies the location on the surface element (39) as an ID of the surface element (39) and coordinates (U, V) in a coordinate system of the surface element (39) and for a the current position specification ( 18) receiving control unit (13, 21) the location is converted into absolute coordinates of the three-dimensional space (24) according to the internal pose specification and a driving trajectory for a automated driving (11) on the driving surface (12) is planned and the motor vehicle (10) is guided on the driving surface (12) by means of driving actuators (15). A method according to any one of the preceding claims, wherein the state space provides a fiber in addition to the 2D position on the manifold (35), the fiber: • a dimension for a yaw angle specification or • two dimensions for a yaw angle specification and an additional roll angle specification or pitch angle specification or • has three dimensions for three solid angle specifications. Method according to one of the preceding claims, wherein the state space is an indication of the state of a chassis of the motor vehicle (10) includes, which is signaled by at least one chassis sensor. Processor circuit (20) for a motor vehicle (10), the processor circuit (20) being set up to carry out a method according to one of the preceding claims. Motor vehicle (10) with at least one sensor (27) and with a processor circuit (20). Claim 13 , which is coupled to the at least one sensor (27).

Images