EP3824247A1

EP3824247A1 - Method and system for determining a position of a vehicle

Info

Publication number: EP3824247A1
Application number: EP19749243.2A
Authority: EP
Inventors: Daniel WILBERS; Christian MERFELS; Bernd Rech; Thilo Schaper; Niklas KOCH; Stefan Jürgens; David Perdomo Lopez; Constanze HUNGAR
Original assignee: MAN Truck and Bus SE; Volkswagen AG
Current assignee: MAN Truck and Bus SE; Volkswagen AG
Priority date: 2018-07-20
Filing date: 2019-07-18
Publication date: 2021-05-26
Also published as: DE102018117660A1; CN113330279A; WO2020016385A1

Abstract

The invention relates to a method for determining the position of a vehicle (1), in which preliminary position data of the vehicle (1) and environmental data are detected, by using the environmental data, landmark measuring data for detected landmarks in an environment of the vehicle (1) are determined, and map data is detected, wherein the map data comprises prior landmark data. A factor graph is determined, wherein the factor graph comprises vehicle position nodes (p0, p1, p2, p3, p4, p5, p6) which represent the preliminary vehicle position data, and landmark position nodes (10, 11), which represent landmark position data, and optimisation of the factor graph is carried out, wherein optimised vehicle position data is determined. The position is determined by using the optimised vehicle position data, . The invention further relates to a system for determining a position of a vehicle (1), comprising a detection unit (2), a processing unit (3), an interface (6) and a computing unit (4). The computing unit (4) is configured to carry out the method according to the invention.

Description

description

Method and system for determining a position of a vehicle

The present invention relates to a method and a system for determining a position of a vehicle.

Many automatic driving functions in modern vehicles require an accurate estimate of the current vehicle pose. Various approaches have been developed in the past for this problem of localization, including, for example, positioning using global navigation satellite systems (GNSS), such as the global positioning system GPS (global positioning system). However, the accuracy of such systems is typically not sufficient for use in automatic driving functions. However, alternative systems often require an excessively high one

Computing effort and therefore - with the computing power typically available in vehicles - too long computing times for control in real time.

DE 10 2013 208 521 A1 describes a method for the collective learning of a highly precise digital street model. A route is traveled several times and trajectory and perception data are recorded. The trajectories are associated and an information graph is formed and optimized, whereby optimal trajectory points are determined. Based on the perception data, a highly precise street model is created on this basis.

DE 10 2016 205 193 A1 describes a method for marginalizing a Poznan graph, in which graphs that get larger and larger over time are reduced. Optimization is carried out to obtain an estimated position. Then a node of the graph is marked and removed. For this purpose, a fixed node is determined, for example a position that has not been changed by the optimization.

In the method described in DE 10 2015 214 338 A1 for determining a pose of a vehicle, a local lane arrangement is recorded relative to the vehicle. Based on a geometric similarity of this recorded lane arrangement

The pose can be updated using the specified map material. An arrangement and a method for sensor fusion are proposed in DE 10 2014 209 340 A1. A factor graph is generated on the basis of received sensor data and a fused sensor signal is generated using an interference machine.

In the method proposed in DE 10 2015 218 041 A1 for the provision of data for a geometry map, odometry data and possibly further data are used

Information such as landmarks used. A maximum travel length is determined as a function of the accuracy of detection for the odometry data.

Finally, in the method described in US 2017/0004379 A1, key points are recorded using images from a vehicle camera and corresponding poses are determined. These poses are used to determine the movement of a vehicle.

The present invention is based on the object of providing a method and a system of the type mentioned at the outset which enable position determination with high accuracy and with efficient use of the resources available in a vehicle.

According to the invention, this object is achieved by a method having the features of claim 1 and a system having the features of claim 9. Advantageous refinements and developments result from the dependent claims.

In the method according to the invention, provisional position data of the vehicle and environmental data are recorded. Landmark measurement data for detected landmarks in an environment of the vehicle are determined on the basis of the environmental data. Map data is acquired, the map data comprising prior landmark data. A factor graph is determined, the factor graph comprising vehicle position nodes that represent preliminary vehicle position data and landmark position nodes that represent landmark position data. The factor graph is optimized, with optimized vehicle position data being determined. The position is determined on the basis of the optimized vehicle position data. This can then be output.

According to the invention, the determination of the position or pose becomes one

Vehicle treated as an optimization problem. The use of factor graphs according to the invention advantageously makes it particularly intuitive and descriptive Description of the problem on which the optimization is based, which in particular represents a system of equations. The optimization problem represented thereby is solved by optimizing the factor graph, whereby this solution can be represented and solved particularly efficiently.

In the process, a factor graph is determined and used to create a

Optimization problem with the least squares method with nonlinear

Represent model functions (nonlinear least squares). As an alternative or in addition, such an optimization problem can be represented directly. The factor graph is used to present the problem in a particularly clear and simple form, which allows simple and safe maintenance and, if necessary, troubleshooting. In addition, particularly good adaptations for fast and efficient processing by a computer are possible.

In the method, a relationship between the accuracy of the localization and the required computing time can also be adapted and, if necessary, adapted according to the specific situation or configuration of the vehicle. This can be done in particular by taking into account a suitable number of measurements from the past and / or by setting a length of the trajectory that is used for localization. The use of a sliding window and the consideration of a large number of positions along a trajectory is discussed in more detail below.

The method also relates to a map-relative localization, so that the specification of the global coordinates in the map data used is less critical than, for example, with known RTK-GPS methods (real-time kinematic). The creation of the maps is therefore less complex and less expensive. Such a map-relative localization can be used in particular for automated driving functions, it also being possible to load further information from other parts of the map if required. If, however, the map is correctly globally referenced, a global localization can be carried out directly on the basis of the map-relative localization according to the invention, ie correct data about the position or the pose are generated in a global coordinate system.

Compared to the other methods according to the prior art, in which the

If raw data acquired by sensors are used, in particular without detection of semantic landmarks in the vicinity of the vehicle, map data with a smaller file size are required, so that advantages in the use of storage space and possibly bandwidth when reloading data are achieved. Furthermore, the Computing time can be shortened. Maps with semantic objects and landmarks can also be checked and serviced particularly easily, as the corresponding ones

Landmarks can be easily found and checked.

When using landmarks, an abstraction layer is used between the raw sensor data and the localization step, and different sensors can be used to detect landmarks. In this way, sensors can be exchanged and changed more easily, for example when changing between different ones

Product generations. In addition, maps with landmarks can be created particularly easily by third-party providers who do not need to know exactly which raw sensor data is available in the vehicle. In contrast, known methods typically assume that features and characteristics are specified in the way they are observed, for example, with a specific camera or a specific type of radar sensor.

In the process, the relevant input data are displayed in a factor graph, this is optimized and an optimized vehicle position can then be read out. The optimization method of the invention also leads to the fact that the

Input prior existing landmark data can also be optimized. This means that, depending on the data collected, these sizes can also be adjusted so that reality is as accurate and accurate as possible.

The use of an optimization method has several advantages over known particle filter methods or EKF (extended Kalman filter): Optimization methods are usually more precise than simple filter methods, which have to allow greater noise in the system in order to deliver non-divergent results. Optimization methods like the one used here are deterministic, which is not the case with a Monte Carlo method like the particle filter. However, non-deterministic methods are fundamentally critically evaluated in questions of functional safety and it therefore seems unlikely that they can form the basis for highly automated driving functions. In addition, non-deterministic methods provide poorly reproducible results because they depend on

Are random sizes. This complicates their maintenance and, if necessary, finding errors. Furthermore, the results of non-deterministic methods are not always identical, which means that the quality of the determined position can fluctuate with the same input variables. With an optimization process, however, limits can be specified in advance depending on its input variables for how well the process will work. In the method according to the invention, positions can be specified in any coordinate systems in a manner known per se. In particular is a global one

Coordinate system is provided or a relative coordinate system, for example relative to the vehicle or another reference point. For example, a coordinate system can be specified within a certain area, for example in a parking lot, in a parking garage or on a private site.

According to the invention, the term “position” should be broadly defined and in particular also include poses. In addition to the position within a coordinate system, these also indicate an orientation, in particular a two-dimensional or three-dimensional orientation relative to the coordinate system. Poses are of particular importance for applications in the field of automotive technology, since they usually also dictate a direction of movement of the vehicle.

Position data therefore include information on the position, possibly also in a broader sense information on the pose of the vehicle and / or on a state of motion of the vehicle. This information can relate to a current or earlier point in time, in particular also a series of successive points in time. The position data are recorded or made available in a manner known per se.

In an embodiment of the method according to the invention, the recorded provisional position data comprise a global position estimate and a local position estimate. As a result, different aspects of the position determination can advantageously be taken into account.

The terminology “local” and “global” describes the coordinate system in particular: local pose and landmark factors are described in the vehicle coordinate system, while global pose and landmark factors are described in a global (world coordinate system).

A global position estimate can be made, for example, using a global

Navigation satellite system such as GPS can be determined. The result of a corresponding GPS measurement can include information about a position and / or a pose of the vehicle. Other methods known per se can alternatively or additionally be used. When optimizing the factor graph, this preliminary information becomes

Vehicle position taken into account so that the vehicle pose to be determined is close to this GPS measurement should lie, but this requirement is only approximately met due to other factors in the graph. It is therefore a prior value that is known at the beginning of the method and is used as the basis for the optimization. A global position estimate can be used in particular for initialization.

A local position estimate can be determined, for example, using odometry methods, for example by means of tire odometry, visual odometry or other methods known per se, such as LIDAR scan / Wafc / 7 / ^' n or inertial navigation. In particular, a shift relative to the vehicle is detected and quantified for a coordinate system, whereby the self-movement can be measured.

The acquisition of the environmental data is also carried out in a manner known per se, in particular in such a way that the data provides information about landmarks and other features in one

Environment of the vehicle can be determined, in particular information about the relative positions of landmarks relative to the vehicle.

In a further embodiment of the method, the environmental data are recorded using a camera, a laser sensor and / or a radar sensor. Optionally, an ultrasonic sensor or an infrared camera can be used alternatively or additionally. Furthermore, further known detector modules or sensors of the vehicle can be used or data can be received from an external source or detection device. As a result, sensors can advantageously be used which are already widespread and inexpensively available in many modern vehicles.

The position and environment data can be recorded in particular for a large number of times and / or positions during the movement of the vehicle. In this case, the movement of the vehicle along a trajectory, that is, along a specific path as a function of time, can be tracked.

The method according to the invention is card-based, which means that it requires that

Map data is present that includes prior landmark data. With similar known localization methods, this data is neglected, which is disadvantageous for the

Accuracy of localization results can result. The card data can be on

are recorded in different ways. For example, they can be provided by a storage unit of the vehicle, such as can be included in a navigation system. Alternatively or in addition, they can be accessed and accessed from an external facility can be received, for example via a computer network such as the Internet or from a local service, for example a server, which provides the card data for a parking lot or an operating site. Furthermore, a transmission through a network of vehicles can be provided. The map data can be recorded in particular when driving through a specific geographical area, for example by providing it when the vehicle is on a specific section of a road, a parking lot or a

Company premises. The acquisition is carried out in particular on the basis of an initial global position estimate, for example using GPS. The card data can also be recorded on the instruction of a user, for example when manually calling up updated map data.

The prior landmark data included in the map data are treated in particular as global landmark factors. They include information about landmarks in a geographic area, in particular their position, orientation and / or other features that can be used for their identification and localization.

“Landmarks” in the sense of the invention are, in a broad sense, features in a geographical environment to which at least one position can be assigned.

In particular, such features are understood as landmarks that can be recorded, identified and evaluated as a pattern using a pattern recognition method. Landmarks can be, for example, road markings and strips, posts, piles and pillars, signs, corners, edges and surfaces of buildings, elements of vegetation, artificial landmarks such as barcodes or other features.

In the method according to the invention, landmark data are determined on the basis of the surrounding data. Landmarks are detected and for the detected ones

Landmarks are determined further information, such as their position, orientation, characteristics, certain characteristics such as color, length, radius or other. The parameters used here depend, in particular, on the type of landmark: for example, a start and an end position can be assigned to a road marking, which gives a length and direction, and also, if appropriate, a thickness, color, shape and / or pattern. In a similar way, the environmental data can be evaluated for further landmarks and other types of landmarks.

In particular, semantic landmarks are assumed, which are based on the

Environment data can be recognized and classified as objects. In this case represents a Landmarks not only a pattern recognizable in the surrounding data, but a type of landmark is recognized, to which certain properties can be assigned. In this case, the raw data are not evaluated directly, for example on the basis of a pattern recognition in the data recorded by a laser scanner, but a landmark is recognized as an object that belongs to a certain category of landmarks and has certain, clearly defined properties that are associated with different methods can be measured. One advantage of this procedure is that the information stored in a map about such semantic landmarks can be checked for correctness particularly easily, in particular also manually. It is also possible to use map data with prior landmark data from any source, making it particularly high

Flexibility is achieved. Because the semantic description is independent of the specific vehicle sensor system, such a map can be used flexibly on a wide variety of vehicles and the prior landmark data can be evaluated independently of the sensor system used.

The factor graph determined in the method consists of a set of nodes and edges, which represent factors here. Here, the nodes comprise estimated values that are to be determined, in particular vehicle poses at specific times and landmarks with their corresponding poses. Vehicle poses are parameterized as points in one

Coordinate system and as an orientation within the coordinate system

(two-dimensional: x, y, theta; three-dimensional: x, y, z, roll, pitch, yaw angle). In particular, a global coordinate system is provided. The parameterization of landmark nodes differs depending on the landmark type, whereby landmarks are divided into classes of different geometries. Landmarks whose position can be represented by a point, such as posts, are parameterized as a point (two-dimensional: x, y; three-dimensional: x, y, z) in a coordinate system. For other types of landmarks, at least one position is also provided as a parameter, possibly together with other parameters.

The edges or factors connect the nodes. A factor can be connected to any number of nodes. In particular, the factors are connected to only one node (unary factors) or they connect two nodes in pairs (binary factors). Mathematically, unary factors describe a prior, i.e. information that is already known, about the linked estimate. Are in the process

in particular two unary factors of central importance: so-called global pose factors and global landmark factors. Global pose factors are connected to a node that represents a vehicle pose and describe the (mathematical) prior about which position or pose this node corresponds to, which is the estimated value for the vehicle pose at the node. This unary factor is formed by a GPS measurement and assigned to a vehicle pose. In the resulting optimization problem, the boundary condition is that the estimate should remain close to the GPS measurement.

Global landmark factors similarly describe a prior over one

Landmark node that represents the parameters of a landmark to be estimated. In the method, they are formed directly from the prior landmark data included in the map data. They are then assigned to a landmark node and linked to it. With the resulting optimization problem, this results in the boundary condition that the estimate should be close to the prior landmark data.

In a further development of the method, when determining the factor graph, a

Association of landmark measurement data carried out, wherein associated landmark measurement data are determined depending on a specific point in time. In particular, several repeated observations of the same landmark are identified in the method. This advantageously makes it possible to reduce the scope of the optimization problem to be solved arithmetically and to solve it more efficiently by repeating

Observations of the same landmark are associated. After the optimization problem has been solved, the results obtained can also be compared with the prior information in order to check the correctness and consistency of the prior information.

Landmark measurement data can be associated with the map data, in particular the prior landmark data. This can be done at any time step for which the method is carried out. As a result, registered landmarks can be assigned to landmarks included in the map data, and incorrect assignments can be corrected if necessary, even retrospectively for a previous time step. The subsequent checking of an assignment or the subsequent correction of an incorrect assignment is made possible in particular by the fact that the assignment can be repeated as a separate method step for each point in time of the landmark data recorded.

The method can also include a history of assignments between local landmark factors, which are determined on the basis of the landmark measurement data, and map landmarks, determined from the prior landmark data. For example, it counts how often a local landmark is assigned to a map landmark. If this number of matches of the mappings reaches or exceeds a threshold, this hypothesis is assumed to be valid. This makes it possible to avoid incorrect assignments. Several association hypotheses can be compared, for example for several points in time for which the method is carried out, and incorrect or defective associations can be prevented or corrected. The association can also be subsequently corrected. Landmarks detected locally by the vehicle and determined on the basis of the prior landmark data can thus be associated with particular certainty.

Based on the history, a local landmark, which was determined on the basis of the landmark measurement data and to which no map landmark could be assigned on the basis of the prior landmark data, can subsequently be a probable map landmark for the

Time of the acquisition of the landmark measurement data can be assigned. Furthermore, incorrect assignments or associations with a too low probability can subsequently be identified on the basis of the history.

In general, it cannot be assumed with the method that the acquisition of the environmental data and the position data always takes place synchronously with one another and exactly at the point in time for which a position is to be estimated. Different methods can therefore be used in order to be able to use the measurements for a specific point in time: The time offset can be ignored, in particular if the times are very short or if it is known that the movement of the vehicle is so slow that there is no relevant shift in position for the time offset. Alternatively, the position data, in particular data of an odometry method, can be used to extrapolate the displacement of the position during the time offset from the movement sequence or to interpolate between two points in time.

During an apprenticeship, the factor graph is determined for points in time that are regular

Intervals follow one another. In particular, this is the point in time for everyone

Factor graph completely redesigned. As a result, a movement of the vehicle along a trajectory can advantageously be tracked. The regular intervals can relate to fixed time intervals. Furthermore, the factor graph can be generated at regular spatial distances or it can be a minimum spatial distance

Condition for the creation of a new factor graph can be provided. Before the Determining a new factor graph can also be checked whether there is data from a particular input source or from several input sources. In particular, at the times when the factor graph is newly determined, new estimates for the position, that is to say new position nodes, are generated.

In a further embodiment, map measurement data and prior landmark data are assigned and the vehicle position data and landmark position data of the factor graph are determined on the basis of the assignment. The assignment is made locally in particular. As a result, associated detections of landmarks and information about the landmarks comprised by the map data are advantageously identified.

The assignment can be based on the information stored on the landmarks

or these detected parameters. In particular, a local proximity between a detected landmark and the predefined estimated position is used for the assignment, that is to say that a detected landmark is assigned to a landmark within the map data on the basis of the landmark measurement data, that of the respective one

Position estimate corresponds as closely as possible.

With a further training in the optimization of the factor graph for the

Landmark position node of the factor graph determines optimized landmark position data. This advantageously also optimizes the positions of the landmarks when optimizing the factor graph. The method can be used to improve the prior landmark data given by the map data on the basis of measurements.

In particular, it can be provided that quality data is generated and output on the basis of the prior landmark data and the optimized landmark position data. This advantageously allows the quality of a detection of landmarks or a quality of map data to be checked. For example, the quality data can include information as to whether and to what extent the landmarks actually detected differ from the information that is provided with the map data. In particular, it can be provided that the map data is updated on the basis of the quality data.

In the case of an apprenticeship, the method is repeated iteratively for a plurality of points in time and a trajectory of positions is output, with a later factor graph each representing one is assigned later, is determined on the basis of an earlier factor graph which is assigned to an earlier time. The factor graph is marginalized in such a way that the later factor graph does not reach a predetermined maximum size

exceeds. In particular, the method is carried out at regular time intervals. This advantageously allows a trajectory of the vehicle to be traced.

The maximum size of the factor graph can be defined in a manner known per se, in particular as the number of nodes and edges and / or on the basis of a complexity measure which quantifies the computation effort required to solve the optimization problem.

It is envisaged that a "sliding window" is used, which defines a fixed time window based on a current point in time. For example, all the data of the factor graph can always be taken into account for a specific time interval before the current time. Older data can be completely ignored, especially if the time interval is long and the older data is therefore out of date. Weighting can also be carried out depending on the age of the data, with older data in particular being weighted less strongly. In addition, the data from earlier times

are summarized, for example, into a single estimate for a parameter for a past period, while more recent data than individual measurements

be taken into account.

In particular, the method involves a completely new optimization of the factor graph for each time step. In contrast, older estimates are typically not improved in conventional methods, since errors in the estimate are continued at earlier points in time. The method therefore advantageously optimizes the estimates at any point in time instead of being dependent on the results of past estimates.

In the method according to the invention, the factor graph comprises landmark position nodes and prior landmark data, which can be represented in particular by global landmark factors. In prior art methods, the landmark position nodes are often removed by marginalization. Marginalizing the

In this case, landmark position nodes can lead to errors that result from the approximation (marginalization errors). This type of error is avoided in the method according to the invention. In the method according to the invention, namely

advantageously receive improved information about the landmark positions. This can be used to check the structure of the factor graph over time for meaningfulness and internal consistency. If necessary, a card quality can also be assessed.

The system according to the invention for determining a position of a vehicle comprises a detection unit for detecting preliminary position data of the vehicle and

Environment data, a processing unit for determining landmark measurement data for detected landmarks in an environment of the vehicle based on the environment data and an interface for receiving map data, the map data comprising prior landmark data. It also includes an arithmetic unit that is set up to determine a factor graph, the factor graph comprising vehicle position nodes that represent preliminary vehicle position data and landmark position nodes that represent landmark position data. It is also set up to optimize the factor graph, determining optimized vehicle position data, and determining and outputting the position on the basis of the optimized vehicle position data.

The system according to the invention is particularly designed to implement the method according to the invention described above. The system thus has the same advantages as the method according to the invention.

The interface for receiving card data can be designed in different ways known per se. For example, an interface to a computer network, such as the Internet, or a local service, such as a local server, can be used to provide map data for a particular environment. As an alternative or in addition, the vehicle can comprise a storage unit or a storage unit of an external device can be used, in particular data from a navigation device being accessible.

In one embodiment of the system according to the invention, the detection unit comprises a camera, an ultrasound sensor, a laser sensor and / or a radar sensor. Alternatively or additionally, other sensors and sensor types can be used. As a result, sensors which are widely used and available in many vehicles are advantageously used, the system not being limited to these sensors. The invention will now be explained using exemplary embodiments with reference to the drawings.

Figure 1 shows a vehicle with an embodiment of the invention

Systems,

FIG. 2A shows an exemplary embodiment of the factor graph according to the invention,

Figure 2B shows an embodiment of a situation on a road with the

associated map section and

Figure 3 shows an embodiment of the method according to the invention.

Exemplary embodiment of the system according to the invention is explained with reference to FIG. 1.

A vehicle 1 comprises a detection unit 2, which is coupled to a control unit 8. The detection unit 2 comprises a plurality of detection and not shown

Sensor modules, including modules for determining the position using GPS (global positioning system) and using odometry, in particular by detecting the wheel speed and the steering angle of the vehicle 1. It also includes a laser scanner and a camera for detecting the surroundings of the vehicle 1. The control unit 8 comprises one

Processing unit 3, a computing unit 4 and a navigation system 5. As an alternative or in addition to the navigation system 5, an interface to a storage system of the vehicle 1 or to an external storage system can be provided. The vehicle 1 also includes drive and steering means 7, which are designed in a manner known per se and allow longitudinal and / or lateral control of the vehicle 1. The vehicle 1 further comprises an interface 6, which is coupled to the control unit 8 and can establish a separable, wireless data connection to an external server 10 in a manner known per se.

An embodiment of the factor graph according to the invention is explained with reference to FIGS. 2A and 2B. It is explained by the above with reference to Figure 1

Embodiment of the system according to the invention.

The factor graph comprises nodes and edges, the edges being referred to as factors in the exemplary embodiment. Here, unary and binary factors are provided, which are connected to one or two nodes. The nodes are shown as circles in FIG. 2A, the factors are represented by rectangles and linear connections to the nodes. The nodes p ₀ to p ₆ represent (vehicle) pose nodes, that is to say the poses (positions and orientations) of the vehicle 1 at specific times. In the exemplary embodiment, the method essentially serves to specify and optimize estimates for these pose nodes, in particular using a global coordinate system.

The nodes l ₀ and h represent landmark nodes, that is, poses of different landmarks.

The local pose factors o ₀ to o ₅ represent odometry factors. They therefore depend on the data of an odometry device measured by the detection unit 3, which during the movement from one pose to another data on the movement,

in particular speed and steering angle, recorded and recorded. In the example, approximately o _{0 represents} the change in the pose of the vehicle 1 between the pose nodes p ₀ and pi.

The global pose factors a ₀ and ai represent prior information about a vehicle pose, the exemplary embodiment here providing for a global position determination by means of the GPS module of the detection unit 3.

The local landmark factors g ₀ to g ₃ represent observations of a landmark from a vehicle pose to a landmark. That is, while the vehicle 1 is in a certain pose, it detects its surroundings with at least the landmarks by means of the detection unit 3. The measurements go into the local landmark factors g ₀ to g ₃ and allow the determination of a relative position of the landmarks relative to the vehicle 1. In the example shown, the landmark belonging to the landmark position node l ₀ was detected from the vehicle positions pi and p ₃ , the landmark at h from p ₃ and p ₅ .

The global landmark factors m ₀ and mi represent prior landmark data. In the exemplary embodiment, these are comprised of map data, in particular absolute positions, orientations and further parameters of different landmarks.

In the case shown in FIG. 2B, there is a vehicle 24 on a road 21 on which road markings 22 are attached and guide posts 23 are arranged on the edge thereof. The vehicle 24 detects its surroundings by means of a detection unit 3, which in the Is essentially designed as shown in Figure 1 for the embodiment of the system according to the invention. Dashed lines indicate that some of the guide posts 23 are detected by the vehicle 24, in particular their positions relative to the position of the vehicle 24 being detected.

Also shown is a section of a map 30 with a map of a road 31, a road marking 32 and a guide post 33. When determining the position according to the method according to the invention, the data on the map 30 together with the detections of the vehicle 24 flow into a factor graph which it Essentially as shown in FIG. 2A, the pose of the vehicle 24 at certain times is represented by the pose nodes p ₀ to p ₆ , the global pose factors a ₀ and ai represent GPS measurements at certain times and the local pose factors o ₀ to o ₅ correspond to a movement of the vehicle 24 between two times in each case. The positions of the landmarks 23 are represented by the landmark nodes l ₀ and h, global landmark factors m ₀ and mi represent prior landmark data as included in the map 30, and the local landmark factors g ₀ to g ₃ represent observations of landmarks 23.

An exemplary embodiment of the method according to the invention is explained with reference to FIG. In this case, the exemplary embodiment of the system according to the invention shown above with reference to FIG. 1 and the one shown in FIGS. 2A and 2B

Embodiments of the factor graph and a situation on a street with the associated map section.

In a first step S10, provisional position data of the vehicle and

Ambient data as well as map data recorded. The map data can be provided by the navigation system 5 or a storage system, for example. These input data can be made available in different ways, in particular by means of the detection unit 2 and via the interface 6 of the vehicle 1.

The input data are buffered in further steps S21, S22, S23, S24, that is, a memory is provided which collects and makes available input data at least until it is processed.

The input data include provisional position data, which in the exemplary embodiment is carried out using an odometry method and as global position estimates, in particular using GPS, be recorded. Global position estimates are buffered in step S21, and odometry data are buffered in step S22.

Based on the environmental data acquired by the acquisition unit 2, the

Processing unit 3 initially detected landmarks in a manner known per se.

In particular, semantic landmarks are determined, that is to say the landmarks are assigned to specific types and parameters for the individual landmarks are determined. The determined parameters include at least one position, possibly further properties of the detected landmark, such as a length or other extent, a radius, a geometric shape or a start and end point. The landmark detection is buffered in step S23.

The map data acquired at the beginning include, in particular, a map with prior landmark data, that is to say information about landmarks in the map

covered area. The map data can be fully loaded at the beginning, reloaded in sections, or from a backend or other vehicles,

in particular by means of vehicle-to-vehicle communication (vehicle-to-vehicle / v2v). The map data, in particular a landmark map, is buffered in step S24.

There can be several sources for the various input data recorded at the beginning, but in the exemplary embodiment at least one source is provided for each type.

In particular, several modules for the acquisition of position data are taken into account and thus large parts of the properties of a location fusion are integrated into the method. If in the following there is talk of one source per input data type, it is assumed that with several sources the respective method steps per source are carried out separately.

The buffering S21, S22, S23, S24 of the input data also takes place concurrently during a cycle, in particular in order to be able to provide data for pose determination at later times. At certain points in time, the graph structure is started and then the graph optimization. Finally, the final pose estimate is extracted from the graph by optimization.

In order to be able to determine a factor graph, the times at which the vehicle pose is to be estimated are first determined. This is done in the embodiment in a fixed time interval, but could also happen, for example, by a spatial minimum distance that the vehicle travels between two estimates. Another possibility is to link these times to the existence of the data of one or more input sources. At these times, Poznan nodes, in particular vehicle position nodes, are then inserted in the graph and a new one can be created

Post estimation done.

Subsequently, in a step S41, global pose factors a ₀ , ai are calculated using the global pose estimates that are present in the buffer. These must fit in time with the pose nodes previously created. Since in general the global post estimates are not recorded at the point in time for which the factor graph is to be determined, the global post estimates must be postponed using a suitable strategy. In the exemplary embodiment, a nearest neighbor strategy is used for this, in which the next point in time is selected. This works well in cases when the time difference between Poznan nodes p ₀ to p _{6 is} small, about only a few milliseconds. An alternative strategy is to interpolate between two global position estimates in the buffer in order to obtain a new estimate that fits in time to a pose node p ₀ to p ₆ . Another strategy is to determine the shift in global posture estimates using odometry. This works well when odometry makes small mistakes for short sections. After determining the global pose factors a ₀ , ai, these are inserted into the graph and connected to the corresponding pose nodes p ₀ to p ₆ .

In a step S42, the odometry data are processed in the buffer. Here too there is the problem that the times of the input data, that is to say the odometry measurements detected by the detection unit 2, generally do not match the times of the pose nodes Po to p ₆ . Here, too, interpolation and closest neighbor processes can be used. In the exemplary embodiment, the odometry measurements are preferably chained to one another by the distance traveled between any points in time

To get vehicle coordinate system, which is ultimately similar to an interpolation. For short periods, an extrapolation of the data can alternatively or additionally be provided. Based on the odometry data processed in this way, the corresponding local pose factors o _o to o ₅ are inserted in the factor graph. In particular, the local pose factors o _o to o ₅ each link two vehicle poses p _o to p ₆ . Either the pose nodes p ₀ to p ₆ that are closest in time to the measurement or two always successive pose nodes p ₀ to p ₆ are preferably linked. The landmark detections buffered in step S23 are also processed in a similar manner. The landmarks are first brought together in a local association step S31. This step S31 does two things: First, the detections are projected onto the time stamp of the closest Poznan node p ₀ to p ₆ . For example, the data from the odometry buffer can be used for this purpose.

Alternatively or additionally, in a previous process cycle, optimized pose nodes p ₀ to p ₆ can be used to determine the movement and to carry out an interpolation or extrapolation. If there is no data in the factor graph at the time the method is carried out, a movement model of the vehicle can be used to extrapolate the movement at a desired point in time.

Furthermore, in step S31, hypotheses are made as to which of the detected landmarks which are stored in the buffer relate to the same physical object in the vicinity of the

Vehicle 1 belong. For example, the same landmark can be detected several times by the same detector or by different sensors. This can be done, for example, using a nearest neighbor strategy, with threshold values for excessive distances leading to rejection of the hypothesis. Another option is to use special descriptors of the

Evaluate landmarks, especially different types of landmarks. In general, a distance function is designed and checked, which of the other possible landmarks is at a shorter distance than a threshold value. If in that

Embodiment found a clear solution, it is accepted as a hypothesis. If there are several possible candidates, the hypothesis formation appears to be too uncertain and the hypothesis is therefore not accepted.

The local association in step S31 thus delivers a set of local landmark factors g ₀ to g ₃ , which are now used directly in a step S43 to determine the factor graph.

In addition, they are processed in a step S32 for a card assignment. The local landmark factors g ₀ to g _{3 are} compared with the map included in the map data and hypotheses are generated which of the local ones

Landmark factors g ₀ to g ₃ agree with which map landmark, that is to say a landmark comprised by the prior landmark data. In step S32 it is thus determined which subset of the map landmarks corresponds to which subset of the local landmark factors g ₀ to g ₃ . The execution of this step S32 can be made more difficult by the fact that map landmarks are missing and / or by incorrect detections of the detector are present, in which case no association can be found in each case.

The procedure is similar to that described above for the case in which hypotheses are formed in order to assign several detections to the same detected landmark. The result of this step S32 is a set of associations between local landmark factors g ₀ to g ₃ and map landmarks. These are then processed in a further step S33 for the time comparison.

In step S33 of the time comparison, a history is formed about the assignments between local landmark factors g ₀ to g ₃ and map landmarks. In which

In this exemplary embodiment, this includes the number of times the hypothesis was generated that a specific local landmark factor factor g ₀ to g ₃ can be assigned to a specific map landmark. This hypothesis is only assumed to be valid after a minimum number of matches, so that incorrect detections can be avoided. In the temporal adjustment step S33, hypotheses for data association are validated over a longer period of time, and short-term bad associations can also be prevented and corrected. In this way, a more stable assignment is achieved between landmarks recorded locally by the vehicle and present on the map.

It may also happen that a local landmark factor g ₀ to g ₃ cannot be assigned to a map landmark in the current step. In this case, the history can still be used to determine on the basis of historical knowledge which assignment is likely or was likely at the time of detection. In addition, incorrect assignments can be found here that are not in line with the history.

Based on the assignments between local landmark factors g ₀ to g ₃ and the

Map landmarks are formed in a step S44 global landmark factors m ₀ , mi. These are then built into the factor graph. In particular, they indicate at which position a particular landmark is expected.

At the time of step S50, the factor graph with the information is one

previous and with additional information of the current time step.

As a result, the function graph has normally become larger in normal operation. In order to avoid excessive enlargement, a sliding window is provided in the exemplary embodiment, that is to say a time window that shifts with each time step. If through If such a shift of information in the function graph falls out of the time window, the function graph must be trimmed accordingly, for which different methods can be used:

In the exemplary embodiment, the oldest nodes and factors are removed outside the time window. Alternatively, only the nodes and factors of the factor graph that have the least information content can be removed. With both strategies, it is important to consider how the "removal" should best be implemented. A first option is to simply delete nodes and factors. This is of low complexity and can be sufficient for a large factor graph. Another option preferred in the exemplary embodiment is the marginalization of information. For this the

System graph of the optimization problem underlying the factor graph with the Schur complement marginalized. This method is known per se, for example from the

Scientific article "Information-theoretical compression of pose graphs for laser-based SLAM", H. Kretzschmar and C. Stachniss, International Journal of Robotics Research, 2012,

Vol. 31, pp. 1219-1230.

In the next step S60, the factor graph is optimized, whereby known algorithms for graph optimization can be used. This optimization provides a determination of the node values, which represent all factors in the best possible way. All pose nodes Po to p ₆ and landmark nodes l _o , h are thus determined so that the conditions of the factors m _o , mi, g _o to g ₃ , o _o to o ₅ , a _o , ai are as good as possible be respected. In particular, the pose nodes p _o to p _{6 are} of interest, those on the graph head are the most recent vehicle poses

represent. In step S70, the most recently determined vehicle pose is usually output. For other applications, it may be of interest to alternatively or additionally output older pose nodes in this step S70, for example in order to carry out calculations for the past. A trajectory that is driven can be relevant for other applications, that is to say a train that is bordered by the vehicle and that can also be output. The determined landmark positions l ₀ , h, which can also be output in this step S70, can be relevant for yet other applications.

In the exemplary embodiment described above, an initial pose is determined by means of GPS measurement or on the basis of comparable data. The factor graph can be built up beforehand, but optimizing it does not bring any meaningful results for most applications, especially if the assignment to the map landmarks is not achieved can. In this case, it can alternatively be provided that the last vehicle pose is saved when the engine is switched off and used for initialization. As a further alternative, a fixed position can be used, which is set, for example, ex works or from a parking space for company vehicles, or which can be configured in a configurable manner, for example a position in a garage for private individuals. In this case, the procedure is independent of GPS measurements and the initial pose can be determined particularly easily.

Additional landmarks can also be integrated in the method if they can be represented geometrically and an error function can be specified for them. This can be implemented for common landmarks using methods known per se, for example for posts, road markings, manhole covers, house surfaces, house edges,

Tunnel entrances, traffic signs, traffic lights, roundabout centers, drains, curbs and similar landmarks.

The global landmark factors m ₀ , mi determined in the method serve in particular as prior landmark data, that is to say as predetermined from the outside of the method

Information about the landmarks of the map, especially information about their positions. As an alternative or in addition, the landmark nodes l ₀ , h of the factor graph can also be marginalized in order to achieve a smaller state vector, which means a reduced memory consumption and a shorter execution time in the construction and optimization of the factor graph. However, losses in accuracy may have to be weighed in favor of speed. Furthermore, when the factor graph is optimized, no estimated landmark positions are obtained on the basis of the landmark nodes l ₀ , h, while the preferred landmark data can also be evaluated in preferred exemplary embodiments of the method.

LIST OF REFERENCE NUMBERS

1 vehicle

2 registration unit

3 processing unit

4 computing unit

5 navigation system

6 interface

7 drive and steering means

8 control unit

10 External server

21 street (real)

22 lane marking (real)

23 guide posts (real)

24 vehicle

30 card

31 street (map)

32 lane marking (map)

33 guide posts (map)

m ₀ , mi global landmark factors

l ₀ , li landmark position node; Landmark nodes

go, gi, g ₂ , g3 local landmark factors

Po, Pi, p ₂ , P3, p ₄ , P _Ö , R _Q vehicle position node; Pose nodes o _o , Oi, o ₂ , o ₃ , o ₄ , o ₅ local pose factors

a _o , ai global pose factors S10 acquisition and preprocessing

521 buffer for global poses

522 buffer for odometry data

523 landmarks (detections)

524 landmarks (map)

531 local association

532 card assignment

533 chronological comparison

541 Determination of global pose factors

542 Determination of local pose factors

543 Determination of local landmark factors

544 Determination of global landmark factors S50 graph structure

S60 graphene optimization

S70 output of position

Claims

claims

1. A method for determining a position of a vehicle (1), in which

provisional position data of the vehicle (1) and environmental data are recorded; Landmark measurement data for detected landmarks in an environment of the vehicle (1) are determined on the basis of the environmental data;

Map data is acquired, the map data including prior landmark data;

a factor graph is determined, the factor graph representing vehicle position nodes (p ₀ , Pi, P2, P3, p ₄ , P _Ö , RQ) representing the provisional vehicle position data, and

Landmark position nodes (l ₀ , h) representing landmark position data;

an optimization of the factor graph is carried out, with optimized vehicle position data being determined; and

the position of the vehicle (1) is determined on the basis of the optimized vehicle position data.

2. The method according to claim 1,

characterized in that

the acquired preliminary position data includes a global position estimate and a local position estimate.

3. The method according to any one of the preceding claims,

characterized in that

when determining the factor graph, an association of landmark measurement data is carried out, with

associated landmark measurement data can be determined depending on a specific point in time.

4. The method according to any one of the preceding claims,

characterized in that

the factor graph is determined for points in time that follow one another at regular intervals.

5. The method according to any one of the preceding claims,

characterized in that

an assignment of landmark measurement data and prior landmark data is carried out and

the vehicle position data and landmark position data of the factor graph are determined on the basis of the assignment.

6. The method according to any one of the preceding claims,

characterized in that

Optimized landmark position data is determined when the factor graph is optimized for the landmark position nodes of the factor graph.

7. The method according to claim 6,

characterized in that

quality data is generated and output based on the prior landmark data and the optimized landmark position data.

8. The method according to any one of the preceding claims,

characterized in that

the process is repeated iteratively for several points in time and a trajectory of positions is output; in which

a later factor graph that is assigned to a later point in time is determined on the basis of an earlier factor graph that is assigned to an earlier point in time; in which

the factor graph is marginalized in such a way that the later factor graph does not exceed a predetermined maximum size.

9. System for determining a position of a vehicle (1), comprising

a detection unit (2) for detecting preliminary position data of the vehicle (1) and ambient data;

a processing unit (3) for determining landmark measurement data for detected landmarks in an environment of the vehicle on the basis of the environment data; an interface (6) for receiving map data, the map data comprising prior landmark data; and

a computing unit (4) which is set up for this: determine a factor graph, the factor graph vehicle position node (p ₀ , ri, p ₂ , p ₃ , p ₄ , Ps, R _Q ) representing preliminary vehicle position data and landmark position node (l ₀ , h), the landmark position data represent includes;

perform an optimization of the factor graph, wherein optimized vehicle position data are determined, and

determine and output the position based on the optimized vehicle position data.

10. System according to claim 9,

characterized in that

the detection unit (2) comprises a camera, an ultrasound sensor, a laser sensor and / or a radar sensor.