DE102019203623A1 - Method of providing a card - Google Patents

Abstract

A method for providing a map (250), sensor data (210) comprising at least several point clouds of an object recognition sensor (12) being available, wherein a comparator unit (32) processes the point clouds of the sensor data (210) and at least one aligned point cloud for a pose provides, wherein reference data (200) comprising ground control points are also available, wherein in a merging unit (34) between groups that are parts of the aligned point cloud and the ground control points, links are established, with an optimization unit (36) the at least one aligned point cloud, the at least a pose, the ground control points, the groups and the links are provided as input data, wherein the optimization unit (36) optimizes the dependencies between the at least one point cloud, the at least one pose, the ground control points, the groups and the links and thereby a map ( 250 In addition, a method for determining position, a computing unit, a memory unit and a motor vehicle are described.

Description

The invention relates to a method for providing a map, a method for position determination, a computer unit, a memory unit and a motor vehicle.

With the ever advancing automation of vehicles, they have to be provided with suitable tools in order to be able to orientate themselves in their surroundings. Commercially available maps that include road networks and GPS information can only provide information to a limited extent for orientation purposes for motor vehicles. In particular, it is advantageous if the cards are provided in a format which is used by the object recognition sensors of the motor vehicle. This enables an efficient comparison between the map and the data obtained in real time.

Autonomous vehicles record their environment through various types of sensors. These can be camera systems, among other things, whereby object recognition sensors such as RADAR, LIDAR and ultrasound are used in particular to locate static and dynamic objects. Object recognition sensors of this type determine targets within a field of vision and mostly determine the relative position and possibly also its speed. These goals determined in real time are used to determine static and dynamic objects and could also be used to determine their position within a map. The goals are usually displayed as a goal list or as a point cloud. The accuracy of the position determination depends, among other things, on the accuracy of the map.

It is therefore an object to provide a map, and in particular a method for providing such a map, which is suitable for determining the position of motor vehicles and their object recognition sensors and also has a high level of accuracy.

The object is achieved by a method for providing a card in accordance with claim 1. Advantageous embodiments of the method are explained in the dependent claims.

Such a map, which is based on target detection by object recognition sensors, is also called a point cloud map.

According to the method for providing a map, sensor data from an object recognition sensor are in particular available in the form of point clouds.

An object recognition sensor is a sensor that detects targets within a field of view and at least determines and outputs their relative position. In particular, an object recognition sensor outputs a target bar or a point cloud with targets. The relative positions of the targets can be provided, for example, in a Cartesian or in a polar coordinate system. The object recognition sensor also provides further information such as a relative speed with particular advantage. At least the relative speed of a target can be determined from the measurement data of the object recognition sensor. Object recognition sensors can be, for example, RADAR, LIDAR or ultrasonic sensors.

Object recognition sensors of this type provide target detections for each measurement, that is to say each measurement cycle. These are mostly provided in the form of point clouds or as target lists. Static as well as dynamic targets and objects can be determined from this data. An object usually comprises several targets that belong to one another, in particular originate from the same object.

The sensor data thus correspond to the point clouds or target lists of several measurements. The goals of these measurements provide the basis for the subsequent map. The map therefore records the areas that have been scanned by an object recognition sensor and its field of vision and are available as sensor data. The map that is provided by the method can include the size of several streets, a city, one or more countries or also continents. The corresponding sensor data is accordingly available as required.

To provide the sensor data, motor vehicles are equipped with these object recognition sensors, which provide the destinations for the areas of the road network to be mapped. The sensor data expediently exclusively comprise static targets.

In a further step, a comparator unit processes the point clouds of the sensor data and provides at least one aligned point cloud for a pose.

The sensor data are provided to the comparator unit. Here, the point clouds present as sensor data are compared with one another and aligned with one another and merged in a common point cloud, which is also referred to as an aligned point cloud. This is also referred to as an aligned point cloud in the following. The comparator unit carries out, for example, the iterative closest point, ICP, the Procedure normal distributions transform, NDT, or a modified or comparable procedure. These methods are well known in the prior art, for example “P. Biber and W. Strasser, "The normal distributions transform: a new approach to laser scan matching," Proceedings 2003 IEEE / RSJ International Conference on Intelligent Robots and Systems (IROS 2003) (Cat. No.03CH37453), Las Vegas, NV, USA, 2003, pp. 2743-2748 vol.3 "or" G. Xu, S. Du and J. Xue, "Precise 2D point set registration using iterative closest algorithm and correntropy," 2016 International Joint Conference on Neural Networks (IJCNN), Vancouver, BC, 2016, pp. 4627-4631 ". The aligned point cloud advantageously comprises only static targets. In the method, inter alia, self-movement data can be used which describes the self-movement of the object recognition sensor or the motor vehicle between two measurements. These serve as an initial estimate for the alignment of two point clouds of successive measurements, which enables the aligned point cloud to be provided more efficiently. The own movement data can be determined, for example, using a suitable acceleration sensor. This optional self-movement data is part of the sensor data.

Such an aligned point cloud advantageously describes a section of the map provided later. In particular, a plurality of aligned point clouds are used to provide the map, and each aligned point cloud is also associated with a pose. The pose provides information for the arrangement of the aligned point clouds with respect to one another. An aligned point cloud accordingly comprises a local coordinate system, which can be converted into a global coordinate system with the help of the pose and vice versa.

The comparator unit thus provides a plurality of aligned point clouds, a pose being present for each aligned point cloud.

In addition to the sensor data, reference data is also available, including ground control points. These ground control points are also called Ground Control Point, GCP. Such ground control points are generally defined and can have different accuracies depending on the method of their determination. The ground control points provided as reference data are advantageously all provided by the same method or by different methods. In the former case, it is only a matter of ground control points of the same type, whereas in the latter case it is a matter of ground control points of different types. In order to provide a particularly accurate map, it is also advantageous to use ground control points that also have a high level of accuracy. In addition, it is advantageous if the ground control points can get within the field of vision of the object recognition sensors. For this purpose, the ground control points are conveniently positioned near the road. Further statements on the ground control points are made in the advantageous design variants of the general part of the description.

In a merging unit, links are established between groups that are part of the aligned point cloud and the ground control points.

The merging unit is provided with the aligned point clouds and the reference data, among other things. Groups of targets are determined from the aligned point clouds and assigned to the ground control points. The group, that is the plurality of targets of the aligned point cloud, and the ground control point are accordingly viewed as the same object, whereby an alignment of the groups and thus also the aligned point cloud can be achieved on the basis of the ground control point. Various criteria can be used in determining whether the group and the ground control point are the same object, such as the shape and dimensions of the object.

In addition to the assignment of groups to ground control points, the merging unit provides further information about the group and the associated ground control point. In particular, this includes a type of quality value that describes how well the group and ground control point match, for example a distance between group and ground control point, their positional accuracy and / or the determined dimensions of the object from group and ground control point. This information is made available for further processing as links in the form of link data. In particular, links are only determined for parts of the ground control points and only for parts of the groups. In other words, not every group is assigned a ground control point, nor is it assigned a group to every ground control point. Further information on the links is provided in the following parts of the description.

The assembly unit can be designed in various ways, which are also explained in more detail in the following parts of the description.

Ground control points advantageously include at least information about a position within the global coordinate system and advantageously information about the accuracy of this position.

The merging unit advantageously determines the links for each aligned point cloud of each pose.

The at least one aligned point cloud, the at least one pose, the ground control points, the groups and the links are fed to an optimization unit as input data, the optimization unit determining the dependencies (links) between the at least one point cloud, the at least one pose, the ground control points, the Groups and the links optimized to provide a map.

The optimization unit records the supplied data and in particular creates a pose graph. This pose graph comprises a plurality of nodes which depend on one another via connections. The Pose Graph becomes a Pose Graph Optimization to optimize the connections and provide a precise map.

Overall, the information of the aligned point clouds and their poses, the ground control points, the groups and the links represent a large number of dependencies. These dependencies can be optimized by adapting their relative alignment between aligned point clouds to ground control points, the aligned point clouds to one another, etc. To provide an accurate map, a pose graph optimization is performed so that the various connections between the nodes are minimized. This provides a global minimum in which the individual point clouds are optimally aligned with one another and also with the ground control points and thus provide a precise map. In particular, the ground control points provide precise landmarks within the global coordinate system, with the aid of which the aligned point clouds of the various poses are aligned. Since each aligned point cloud usually has a plurality of ground control points that can also be linked to different groups, a precise alignment is provided.

The pose graph is explained again in detail below with reference to the description of the figures.

This precise map, in particular point cloud map, enables a precise determination of one's own position to be carried out based on sensor data measured in real time.

The creation of a pose graph and also the pose graph optimization is described in the publications " Kümmerle, Rainer & Grisetti, Giorgio & Strasdat, Hauke & Konolige, Kurt & Burgard, Wolfram. (2011) . G2o: A general framework for graph optimization. Proc. of the IEEE Int. Conf. on Robotics and Automation (ICRA). 3607 - 3613 “.

With regard to this method, the data from the processing from the comparator unit to the merging unit of several sets of sensor data can also be fed to the optimization unit at the same time. The sensor data can come from several sensors of the same type, for example two radar sensors, or also of different types, for example a radar sensor and a lidar sensor. Due to the simultaneous processing of two sets of sensor data from object recognition sensors of different types, in particular point clouds of object recognition sensors of different types can be linked within a single map. The individual units can be used essentially without major changes, that is to say identically, with a different parameterization of these units usually being sufficient to use them for different sensor types.

The link corresponds to the assignment of a ground control point to a group and in particular comprises at least one quality value with regard to the assignment of a group to a ground control point, such as a distance. The reference data include, in particular, the position, positional accuracy and / or dimensions of the ground control points.

Depending on its size, the map provided can consume a particularly large amount of storage space. Such maps are advantageously provided in real time via a radio link, in particular a cellular radio link. In order to reduce the amount of data transmitted, only partial sections of the card are preferably provided. The map or the partial sections of the map are expediently temporarily stored locally in a storage unit. As a result, the position can be determined with high accuracy in real time.

The method explained can be used to update the map regularly, for example daily. This means that a current map is provided at all times. In particular, such a method can be used to automate the acquisition of data. The provision of the sensor data and the reference data can also be automated.

Advantageous variant embodiments of the method are described below.

It is proposed that the merging unit has a grouping unit and an association unit, the grouping unit determining groups of the aligned point cloud and the comparing unit comparing the groups with the reference data in order to determine and link at least some of the groups and at least some of the ground control points to provide.

In this embodiment variant, the merging unit comprises several units. The grouping unit determines groups of targets that correspond to potential ground control points. The ground control points mostly have characteristic features, for example these are rod-shaped. The characteristics of the ground control points are known, whereby the grouping unit examines the targets of the aligned point cloud for these characteristics and provides potential ground control points in the form of groups.

In a particularly advantageous embodiment variant, the ground control points are determined by a synthetic aperture radar, SAR. The SAR is installed in an aircraft and provides ground control points, the position of which is determined with high accuracy within the global map system. Such a SAR detects rod-shaped and metallic objects. The targets of the aligned point cloud are then examined for groups that provide the characteristic bar-shaped structure.

The grouping unit preferably carries out the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) method in order to determine the groups. Alternatively, the groups can be determined by similar or modified methods. To determine the groups, the grouping unit can preferably also be designed as a neural network to which the aligned point cloud is fed as input data. The neural network then outputs the groups.

Within the association unit, the groups are then compared with the reference data which comprise the ground control points. Some of the groups are assigned to a respective ground control point or identified as a ground control point. Accordingly, the group and the ground control point represent the detection of the same object.

This association can take place on the basis of various criteria, for example through spatial proximity and / or dimensions of the detected object and / or also position patterns. The ground control points of the reference data also form certain patterns or relative arrangements which, if necessary, can be found identically or similarly in the groups of the aligned point clouds. The above-described variants of associating a ground control point with a group are only given as examples and are not exhaustive.

A ground control point is advantageously only associated with a single group. In particular, a floor control point is determined for at least a partial number of the groups, so that there is not a floor control point for every group. In most cases, a group is not assigned to each ground control point. Accordingly, the groups and ground control points are assigned to one another and the links are provided.

The link includes, among other things, information about the association of the group with the ground control point and, advantageously, a quality value. The quality value provides, for example, a relative position, a distance between the group and the floor control point and / or a match of the dimensions. These variables are used to solve the pose graph in order to optimize the alignment of the individual nodes with one another.

The groups and the ground control values as well as the links are then made available for further processing.

The association unit uses the Nearest Neighbor method, for example, see also “N. S. Altman (1992) An Introduction to Kernel and Nearest-Neighbor Nonparametric Regression, The American Statistician, 46: 3, 175-185 ", or Joint Compatibility Test, see" Data Association in Stochastic Mapping Using the Joint Compatibility Test, IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 17, NO. 6, DECEMBER 2001 ". Appropriate alternatives or modified forms of these methods can also be used. In a further alternative variant, the comparator unit can also be formed by a neural network to which the ground control points and the groups of aligned point clouds are fed. Accordingly, the neural network outputs the relevant groups, the associated ground control points and also the links so that they are available for further processing.

In a second variant of the merging unit, it has a grouping unit and an association unit, the grouping unit being supplied with the reference data and the at least one aligned point cloud so that groups are determined within the aligned point cloud Correspond to ground control points, and wherein the association unit compares the groups with the reference data in order to provide links between at least some of the groups and at least some of the ground control points.

In this approach, the ground control points are used as a starting point in order to determine the groups and ground control points that belong together by analyzing the objectives of the aligned point clouds. In particular, it is examined whether corresponding targets or clusters of targets were detected in the vicinity of such a ground control point, which correspond to the ground control point. Here too, in addition to the relative position and patterns, the dimensions of the detected dimensions of the object can also be checked.

The grouping unit can be designed, for example, as a neural network. The difference to the variant explained above is that the grouping unit is provided with the reference data in addition to the aligned point cloud, the latter serving as a starting point.

The association unit is designed essentially identically to the embodiment variant described above. In particular, the association unit provides the link.

In a third embodiment variant of the merging unit, this is formed by a neural network. The aligned point clouds and the reference data are fed to the neural network as input data, the neural network outputting pairs of groups of the aligned point cloud and associated ground control points and providing a link for each pair of group and ground control point.

With the third variant of the merging unit, the output data are essentially identical to the other two variants of the merging unit.

It is proposed that the reference data have ground control points which are formed by rod-shaped objects.

As already explained, aerial photographs of an SAR are preferably used for this purpose, which, among other things, provide the dimensions and position of the detected objects with high accuracy as ground control points. Such a SAR detects the position and dimensions of rod-shaped and metallic objects and also provides information about the accuracy of these parameters.

The ground reference points are advantageously provided in a filtered manner so that only the ground control points for which detection by an object recognition sensor of a vehicle is also likely are part of the reference data. This is particularly the case for ground control points that do not exceed a distance from a roadway. For this purpose, for example, the trajectories of streets are used, which are broadened. The reference data then only include the ground control points that lie within the provided corridors of the trajectories. Ground control points that are outside are discarded.

In principle, however, such ground control points can also be provided in a different manner, the accuracy of the provided point cloud map depending on the accuracy of the reference data.

The at least one aligned point cloud, the at least one pose, the ground control points, the groups and the links for at least two object recognition sensors are advantageously determined and these are processed together in the optimization unit.

As a result, the point clouds from object recognition sensors of different types can be displayed and linked within a map, in particular point cloud map. Likewise, several object recognition sensors of the same type can be used for this purpose in order to create an improved and more precise map.

The determination of the input data in the optimization unit takes place for both object recognition sensors preferably independently of one another, the data of both object recognition sensors being processed jointly within the optimization unit in order to provide a single card. For the preprocessing in advance of the optimization unit, the described units can be used for sensor data from object recognition sensors of various types. The processing can for example take place one after the other. In particular, the units can be used unchanged, with mostly only the parameterization of the unit being adapted for the sensor data of object recognition sensors of various types. However, separate units can also be formed for each set of sensor data.

It is further proposed that the object recognition sensors are designed as sensors of different types.

These can be, for example, RADAR, LIDAR or ultrasonic sensors. In particular, a map can thereby be provided which can be used for the precise determination of one's own position on the basis of the measurement data from various object recognition sensors.

The object is also achieved by a method for determining the position according to claim 8.

Accordingly, measured data are determined in real time by an object recognition sensor. In addition, there is a map or parts of a map that was created according to one of the methods according to claims 1 to 7 or at least part of the preceding statements. This is followed by a comparison between the measurement data and the map in order to determine your own position.

This enables a particularly precise determination of the position of motor vehicles. The high accuracy of the map can be converted into an exact position determination with the help of the method. In particular, this method can improve the accuracy of a rough position determination that is provided by another system, such as, for example, a GPS sensor. The map is conveniently available locally on the motor vehicle, for example on a storage medium, or transmitted in real time via a radio link. A combination of data transmission via a radio link with a buffer for a partial map is also advantageous.

The object is also achieved by a computing unit which carries out a method according to one of Claims 1 to 8.

The computing unit includes, in particular, an input interface for receiving sensor data and reference data. In addition, the computing unit includes an intermediate storage unit for the intermediate storage of sensor data and reference data. Furthermore, the computing unit comprises a comparator unit, a merging unit and an optimization unit. In particular, the merging unit forms a neural network or the merging unit comprises a grouping unit and a comparator unit. In addition, the computing unit has an output interface which enables the map to be transferred to a memory unit, the map being stored within the memory unit, preferably a server.

The object is also achieved by a memory unit which has a program code which uses a method according to one of claims 1 to 8 or which stores a map or parts of a map that was created according to one of the methods of claims 1 to 7.

The object is also achieved by a motor vehicle according to claim 11.

Such a motor vehicle uses in particular a method according to claim 8 or is used for a method for providing sensor data according to one of claims 1 to 8.

A motor vehicle that uses the method according to claim 8 can carry out a particularly precise position determination. In addition, motor vehicles with an object recognition sensor can be used to provide sensor data. The acquisition takes place automatically, with the sensor data being advantageously transmitted via a radio link.

These serve to provide a precise and exact map, which also enables an exact position determination.

The term radio connection only refers to data transmission from or to the motor vehicle. It is of course obvious that the data can also be transmitted to a server in other ways after the first transmission link, which is provided by radio. These can be cable lines or fiber optic lines, for example.

• 1 ein Kraftfahrzeug mit Objekterkennungssensor und einer Cloud;
• 2 eine Recheneinheit des Kraftfahrzeugs;
• 3 ein Ablaufplan für die Bereitstellung einer präzisen Karte;
• 4 eine schematische Kartendarstellung zur Visualisierung von Zielen und Bodenkontrollpunkten;
• 5 ein Pose Graph der zur Bereitstellung einer präzisen Karte verwendet wird;
• 6 Verfahren für ein Kraftfahrzeug nach 1 zur präzisen Bestimmung der eigenen Position.

Furthermore, an exemplary and detailed explanation of the method for creating a map, the method for determining the position, the computing unit, the memory unit and a motor vehicle is provided on the basis of several figures. Show it:

• 1 a motor vehicle with an object recognition sensor and a cloud;
• 2 a computing unit of the motor vehicle;
• 3 a schedule for providing an accurate map;
• 4th a schematic map display for the visualization of targets and ground control points;
• 5 a pose graph used to provide an accurate map;
• 6th Method for a motor vehicle according to 1 for precise determination of your own position.

In the 1 is a motor vehicle 10 shown. The vehicle shown schematically 10 is designed here as a partially or fully autonomous vehicle. In particular, this includes an object recognition sensor 12th , a control unit 14th , a storage unit 15th as well as a communication unit 16 . The control unit 14th , by a single or multiple units can be formed is on the one hand with the object recognition sensor 12th and on the other hand with the communication unit 16 connected. In particular, on the motor vehicle 10 also a plurality of object recognition sensors 12th be trained. The further statements are made for a single object sensor as an example.

The object recognition sensor 12th is trained to do objects 18th to be detected in the surroundings of the motor vehicle, in particular within a field of view. These objects 18th can be static or dynamic objects. Examples are in the 1 A tree, a building and a street sign are represented as static objects and a pedestrian as a dynamic object. Such object recognition sensors usually detect the environment by emitting radiation, in particular electromagnetic radiation, with an object 18th reflected radiation from the object recognition sensor 12th is detected again. For such object detection sensors 12th A wide variety of types are known from the prior art.

In particular, it is the object recognition sensor 12th a radar sensor or a lidar sensor. These two types of sensors detect the environment based on the reflections from the objects 18th to be provided. Accordingly, there will only be reflections from objects 18th determined. With regard to the reflections, a relative position is usually determined, the reflection also being referred to as the target. The provision of reflections is usually in the form of target lists, which can also be displayed as point clouds. An object usually provides several reflection points so that an object usually has a large number of targets associated with it. In particular, such an object recognition sensor provides detected targets, at least their position relative to the object recognition sensor being known.

In the case of a radar measuring system, for example, the relative position of a target is provided by a distance as well as an elevation angle and an azimuth angle. In addition, a radar measurement system also provides a relative speed of such a target. Alternatively, the position can also be provided in a different type of coordinate system, with a conversion taking place if necessary. The point clouds determined in real time are generated by the control unit 14th processed to provide information about the surroundings of the motor vehicle 10 to get. In addition, such a point cloud determined in real time can also be used to determine the position and advantageously also the orientation of the motor vehicle within a point cloud map.

Positioning of motor vehicles is provided, among other things, via satellites, in particular GPS or GLONASS. This enables a position to be determined with an accuracy of a few meters. In particular, a D-GPS can also be supported here in order to provide increased accuracy. However, for autonomous driving it is necessary to determine your own position with even greater accuracy.

Appropriate maps will be provided for such a position determination. Such maps are in particular point cloud maps which essentially correspond to the sensor data or point clouds determined in real time. The targets and objects determined in real time in the form of measurement data can be compared with the map material in order to provide a much more precise position determination.

Such a card can for example by a memory unit 15th , which is formed on the motor vehicle, or via a communication unit 16 to be provided. For example, a combination of storage unit and communication unit can also be provided, the storage unit 15th serves as a buffer for map material. The buffer here includes parts of the map, in particular the map parts within a specific area of the motor vehicle.

The communication unit can, for example, be a radio link, in particular a mobile radio link 20th with a server 24 can be provided within a cloud 22nd is located. In this case, only part of the connection can be provided via radio or mobile radio and other parts via cables or fiber optic lines. In particular, it is the server 24 a single server or a plurality of servers connected to one another. On the server 24 the map material is held, preferably on a storage unit 28 . If necessary, in particular at the request of the vehicle, the card or parts of the card are transferred to the motor vehicle. The card or its parts are over the radio link 20th transferred and conveniently in the storage unit 15th saved locally. Current map material is stored in the cloud 22nd kept ready at all times.

In the following, a method is described with which such a map material, that is to say a map, is provided which enables the exact position of the motor vehicle to be determined 10 enables. Accordingly, the map has a plurality of targets with high positional accuracy. This card is made by a computing unit 26th provided, preferably at the server 24 . In particular, the computing unit performs 26th the procedure explained below. Conveniently, the server 24 a storage unit 28 which comprises a program code for the method described below. The arithmetic unit 26th can call up the program code and the data to be processed in order to provide the card.

In the 2 is the arithmetic unit 26th shown in more detail. In addition, the 3 the schedule of the procedure. The arithmetic unit 26th includes an input interface 38 , a comparator unit 32 , a merging unit 34 , an optimization unit 36 as well as an output interface 38 . In addition, the computing unit includes an intermediate storage unit, which is not shown. The merge unit 34 can be designed in various ways and in particular include subunits. In the following, three different variants for the merging unit are explained, these being represented within the figures by appended letters. The reference numbers of variant 1 are appended with an a, variant 2 with a b and variant 3 with a c. This marking is also carried out for the various steps according to 3 .

In a first step 110 the data to be processed are made available via the input interface 30th the arithmetic unit 26th to be provided. The input data include at least the reference data 200 with the ground control points as well as the sensor data 210 which comprise the point clouds. The sensor data 210 include at least the point clouds of a radar object recognition sensor, including radar data 211 called, and / or the point clouds of a lidar object recognition sensor, also lidar data 212 called. In particular, both point clouds 211 and 212 to be provided. The provision of the own movement data 213 is optional.

In the following, the explanations for the point clouds of a radar object recognition sensor are made. The method for point clouds of a lidar object recognition sensor is essentially identical, the explanations applying accordingly. The individual units may have a different parameterization for this purpose. However, this does not change the sequence of the individual process steps.

If the provision of the own movement data 213 not done, this information is also not available for the other blocks, this being indicated by the dashed line 214 is made clear. The use of the own movement data, however, provides a particularly advantageous and efficient determination of the map. In particular, the intrinsic movement data provide the relative movement of the object recognition sensor between different measurement cycles. The radar data 211 represent the corresponding individual measurements of the successive measuring cycles and the point clouds determined therefrom, which were provided for the different positions of the object recognition sensor within the local and global map system.

The reference data include the ground control points, which have precise information about individually defined targets within the map. In this exemplary embodiment, the reference data are provided via a SAR radar, which provides a precise determination of the position of rod-shaped and metallic objects. These can be, for example, posts, poles for traffic signs, traffic lights, lamp posts and so on. The reference data also include information about the shape and dimensions of the corresponding object.

The reference data and the sensor data and in particular their point clouds are in the 4th shown as an example in a schematic map. This shows a lined map, on the one hand the ground control points and on the other hand the targets of the point clouds of the object recognition sensor. The point clouds only provide static targets and the ground control points only provide static objects. To provide the static targets and objects, preprocessing of the raw object recognition sensor data may be necessary.

In the 4th you can see a street from above, which represents an intersection. In the 4th are the targets of the point clouds through crosses 40 shown. In particular, these are the goals of the aligned point clouds, which are explained in more detail below. The ground control points are by squares 42 and circles 44 shown. The circles 44 represent ground control points to which a group of targets are assigned as an identical object. The squares 42 represent ground control points of the reference data to which no group could be assigned. This makes it possible to assign the targets to the individual ground control points, which enables the point clouds or the targets to be aligned with respect to the ground control points.

In a further step 120 are the sensor data first 210 processed. This takes place within the comparator unit 32 holding the radar data 211 and the movement data 213 can be entered. Within the comparator unit 32 the point clouds of the individual measurement cycles are aligned with one another and accumulatively with one another overlaid to provide an aligned point cloud. The individual point clouds are thus combined to form a larger, aligned point cloud. The accumulated aligned point cloud thus covers a spatial area that is larger than that of an individual point cloud of a measurement cycle. In particular, several such aligned point clouds are provided, each aligned point cloud having a pose which describes the position within a global coordinate system. Each aligned point cloud has a local coordinate system, with a conversion from local to global coordinate system and vice versa being possible at any time.

The own movement data is used here more optionally to provide an initial estimate for the alignment of the individual point clouds. This allows the process to run much more efficiently.

The provision of the aligned point clouds takes place, for example, according to the method interative-closest point, ICP, or the method normal distributions transform, NDT or another known approach.

The processing according to step 120 can be done equally for the radar data and lidar data. The comparator unit 32 after performing the step 120 an aligned lidar point cloud or an aligned radar point cloud ready. This is according to the line 220 to the optimization unit 36 and according to the line 224 given to the merging unit.

The merging unit is designed for the aligned point cloud and the reference data 200 to process and provide a link between the ground control points and the aligned point clouds. In particular, targets or groups of targets are identified which correspond to the specific ground control points of the reference data. The assignment of groups to ground control points can, among other things, be determined on the basis of various variables such as reflectivity, dimensions and relative position. In addition, corresponding patterns can also be specified by several ground control points, which are also provided by corresponding groups of the point cloud.

In a first variant, characterized by the letter a, the merging unit is a neural network 34a educated. The neural network 34a becomes the aligned point cloud, here through the line 224 and the reference data represented by the line 228 added, according to step 130a , which is represented by a box with a wavy outline, the processing takes place. The neural network is trained to use the input data to identify the groups of the aligned point clouds that belong to certain ground control points and to provide the associated links for further processing. The link includes at least one quality value that is used for the optimization in a subsequent step. Over the line 230 the relevant groups and the links for further processing in one step 150 provided. The transmission takes place here to the optimization unit 36 .

The merging unit is in another variant 34 by a grouping unit 33b and an association unit 35b educated. In this second variant, the grouping unit first receives the aligned point clouds according to the line 224 fed. In the step 130b groups of targets are then determined, which can in principle be ground control points. For this purpose, appropriate criteria are created to identify these groups. The determination can be carried out, for example, by the generally known DBSCAN method. Alternatively, the grouping unit 33b can also be formed by a neural network, which is only supplied with the aligned point cloud, the groups being output.

After the step 130b That provides the grouping of the goals are these groups in a further step 140b within the association unit 35b processed. The transmission takes place via the line 234 . The association unit 35b are also the reference data 200 according to the line 228 supplied, according to step 140b the association is carried out. Here, the groups are compared with the known ground control points and, if an assignment is made, corresponding quality parameters, i.e. the link, are provided.

The association unit uses in particular the nearest neighbor method or also the joint compatibility test to provide the corresponding links.

The link describes in particular the quality of the assignment of the group to the ground control value, that is to say, among other things, how great the agreement of the dimensions is and how great the distance between the group and the ground control value is. These quality parameters are only examples. The association unit then makes a further step 150 the optimization unit 36 according to the line 230 the groups, the ground control values and also the links are available.

In a third variant c, which also has a grouping unit 33c and an association unit 35c the data are evaluated in a third way.

According to the third variant, to carry out the step 130c the grouping unit in addition to the aligned point clouds according to the lines 224 also the ground control points, i.e. the reference data 200 , according to the line 234 provided. The line 238 is shown in phantom. In the grouping unit, the ground control points are then used as starting points and, based on these, groups are searched for in a corresponding environment around the ground control points. These identified groups and goals are then according to the lines 234 to the association unit 35c transfer. According to step 140c are through the association unit 35c the links are determined. Also according to variant 3 become the data of the optimization unit 36 by step 150 according to line 230 fed.

The steps 130b , 140b , 130a , 130c and 140c can be executed several times if necessary, in order to provide the appropriate groups and links for each aligned point cloud and pose.

According to the step 150 a pose graph optimization is carried out. To do this, a pose graph is first created, which can be found in the 5 is shown. The 5 shows the pose graph 50 for a single object recognition sensor or aligned point clouds of a single object recognition sensor. The pose graph describes various nodes, represented by ellipses and corresponding links between these ellipses. The pose graph is then subjected to a pose graph optimization in order to provide a map with high accuracy.

A central point is provided by the ground control points, i.e. the reference data that is passed through the node 52 are shown. These represent a precise position on the map, especially in the global coordinate system. This ground control point node 52 is about connections 54 with the group nodes 56 connected. The nodes are marked with a .1 to .n to make it clear that they belong to a particular pose. The connections 54 provide the links between the ground control points and the corresponding groups of the respective aligned point clouds and their poses. The group nodes 56 are each with a point cloud node 58 , which represents the individual aligned point clouds, via a connection 60 connected .. The connection 60 takes place between the group node 56.k and the associated point cloud 58.k. In addition, a point cloud 58.k with its respective pose node 62.k is connected via a connection 64 connected. In addition, there are also the poses, i.e. their relative distance between the poses 62 about connections 66 certainly. Also the individual group nodes 56 are interconnected via connections 68 connected.

According to step 150 the connections are optimized. This is done using a cost function, with the costs for the individual connections being minimized. Accordingly, the alignment takes place, among other things, on the basis of the determined quality values. After the optimization, the result is the map 250 based on the sensor data and the reference data. The provision of the card 250 takes place according to the line 240 . The point clouds and the ground control points are optimally aligned with one another and enable precise positioning in real time.

According to the lines 222 , 226 , 232 and 236 can in addition to the set of radar data 211 also a set of lidar data 212 be evaluated. As already mentioned, the structure of the units and the sequence of steps remain essentially unchanged. The data can come from object recognition sensors of the same or different types.

The pose graph looks like in 5 , the same structure being formed again for the set of the second data. The intersection is here the ground control point nodes. In essence, the representation is simply mirrored up with respect to the ground control point node. In this way, point clouds of different types of object recognition sensors can be displayed against one another within a map and also aligned with one another. This enables a significant improvement in the determination of the position, in particular when several object recognition sensors of different types are used, but this enables multiple comparison.

The point cloud map provided by the method is stored in a memory or a storage unit through the output interface 28 of the server 24 stored and is therefore available for retrieval.

In the 6th the individual steps of a method are described, which a motor vehicle 10 performs in order to use an improved position determination based on such a map. This procedure according to 6th can for example by the control unit 14th of the motor vehicle 10 are executed.

In one step 300 targets within the field of view are detected in real time with the help of an object recognition sensor and the control unit 14th to provide. These measurement data recorded in real time represent, in particular, point clouds or target lists. If necessary, the measurement data are preprocessed so that the target list only includes static targets. A measurement point cloud is provided accordingly.

In a further step 310 a card is provided. The provision can be made in real time via a communication unit 16 take place or this is already in a storage unit 15th before, which can also serve as a buffer. In particular, the card or parts of the card are sent to the control unit 14th transferred so that according to step 320 the sensor data can be compared with the map in order to determine the position. The position is determined in real time, whereby the measurement point cloud is compared with the map. In this case, for example, a pattern recognition takes place, which enables the point cloud to be aligned with respect to the map, thereby enabling precise position determination.

Such a computer program, which the method according to the 6th executes, for example in a memory unit 15th of the motor vehicle must be stored so that this can be carried out at any time.

The car 10 can also be designed, among other things, to store corresponding sensor data in its memory unit 15th and via the communication unit 16 to the cloud 22nd on the server 24 , in particular on a storage unit 28 of the server 24 , transferred to. The storage unit 28 of the server includes the program code and the data required to create the map. In particular, the motor vehicle delivers 10 thereby the sensor data with the point clouds and possibly the own movement data. The sensor data can, for example, from a single or several such motor vehicles 10 that are used in daily traffic.

List of reference symbols

10

Motor vehicle

12

Object detection sensor

14th

Control unit

15th

Storage unit

16

Communication unit

18th

Objects

20th

Radio link

22nd

Cloud

24

server

26th

Arithmetic unit

28

Storage unit

30th

Input interface

32

Comparator unit

33b

Grouping unit

333

Grouping unit

34

Merging unit

34a

neural network

35b

Association unit

35c

Association unit

36

Optimization unit

38

Output interface

40

Target (cross)

42

Ground control point (square)

44

Ground control point (circle)

50

Pose graph

52

Nodes (ground control points)

54

Connection (ground control point - group)

56

Nodes (groups)

58

Nodes (aligned point clouds)

60

Connection (aligned point cloud - group)

62

Knot (pose)

64

Connection (Pose - Aligned Point Cloud)

66

Connection (pose - pose)

68

Connection (group - group)

100

step

110

step

120

step

130a

step

130b

step

140

step

150

step

200

Reference data

210

Sensor data

211

Radar data

212

Lidar data

213

Own movement data

214

line

220

line

222

line

224

line

226

line

228

line

230

line

232

line

234

line

236

line

238

line

240

line

250

map

300

step

310

step

320

step

QUOTES INCLUDED IN THE DESCRIPTION

This list of the documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Non-patent literature cited

• Kümmerle, Rainer & Grisetti, Giorgio & Strasdat, Hauke & Konolige, Kurt & Burgard, Wolfram. (2011) [0028]

Claims (11)

Method for providing a card (250), wherein • Sensor data (210) which include at least several point clouds of an object recognition sensor (12) are available, • wherein a comparator unit (32) processes the point clouds of the sensor data (210) and provides at least one aligned point cloud for a pose, • Whereby reference data (200) comprehensive ground control points are available • where links are established in a merging unit (34) between groups which are parts of the aligned point cloud and the ground control points, • the at least one aligned point cloud, the at least one pose, the ground control points, the groups and the links being provided as input data to an optimization unit (36), the optimization unit (36) determining the dependencies between the at least one point cloud, the at least one pose, the ground control points, the groups and the links are optimized and a map (250) is thereby provided. Method according to one of the preceding claims, characterized in that the merging unit (34) has a grouping unit (33) and an association unit (35), wherein the grouping unit (33) determines groups of the aligned point cloud and wherein the association unit (35) identifies the groups compares the reference data in order to determine and provide links between at least some of the groups and at least some of the ground control points. Method according to one of the preceding claims, characterized in that the merging unit (34) has a grouping unit (33) and an association unit (35), the grouping unit (33) being supplied with the reference data and the at least one aligned point cloud so that within the aligned point cloud Groups are determined which correspond to ground control points and wherein the association unit (35) compares the groups with the reference data in order to provide links between at least some of the groups and at least some of the ground control points. Method according to one of the preceding claims, characterized in that the merging unit (34) is formed by a neural network, the aligned point clouds and the reference data being fed to the neural network as input data, the neural network pairs of groups of the aligned point cloud and associated Outputs ground control points and provides a link for each pair of group and ground control point. Method according to one of the preceding claims, characterized in that the reference data (200) have ground control points which are formed by rod-shaped objects. Method according to one of the preceding claims, characterized in that at least the at least one aligned point cloud, the at least one pose, the ground control points, the groups and the links for at least two object recognition sensors are determined and these jointly processed in the optimization unit (36). Method according to the preceding claim, characterized in that the object recognition sensors are designed as sensors of different types. Method for determining position, where measured data are determined in real time by an object recognition sensor (12), where a map (250) or parts of a map are present which are produced according to one of the methods according to Claims 1 to 7th has been created, • A comparison is made between the measurement data and the map (250) in order to determine your own position. Computing unit (26) which has a method according to one of the Claims 1 to 8th performs. A memory unit (15, 28) which has a program code which enables a method according to one of the Claims 1 to 8th uses, or on which a card (250) or parts of a card are stored, which according to one of the methods according to the Claims 1 to 7th was created. Motor vehicle (10), which the method according to Claim 8 uses or for a method for providing sensor data (210) according to one of the Claims 1 to 8th is being used.

Images