METHOD FOR DETERMINING AND PROVIDING TRACK TRACES

Description

[0001] The invention relates to a method for determining and providing lane layouts of roads by means of a central computer unit that is data-technically coupled to vehicles of a vehicle fleet.

[0002] German patent application DE 10 2013 211 696 A1 discloses a method for completing and / or updating a digital road map of a geographical region, a device for a motor vehicle, and a motor vehicle. The method comprises the following steps: a) capturing a digital image of a road segment using a camera of a motor vehicle located on the road segment; b) determining the current position of the motor vehicle in the latitude of the road segment and / or at least one property of the road segment based on the image using a computer in the motor vehicle; c) transmitting a data set to a server containing information about the geographical position of the image capture as well as information about the at least one determined property of the road segment and / or the position in the latitude.Steps a) to c) are carried out by a large number of motor vehicles located in the region, and the digital road map is completed and / or updated using the received data sets via the server facility.

[0003] From DE 10 2013 018 315 A1 a method for providing an environment model for a vehicle is known, in which it is provided to use a grid map for a representation of the environment model and to adapt this grid map to a lane profile, e.g. in the case of a curvature of the lane, the cells of the grid map are also arranged in a curved manner.

[0004] From DE 10 2013 208 521 A1 a method for creating a road model is known, wherein it is provided to determine trajectory data from fleet data and to create a road model from these trajectory data using statistical methods.

[0005] From DE 10 2019 119 002 A1 a method for determining a driving lane is known, wherein it is provided to scan driving lane markings, determine orientations of the scans and cluster the scans taking into account the orientations.

[0006] From EP 2 888 604 B2 a method for determining a lane path for a vehicle is known, wherein it is provided to determine the position of the vehicle in a grid map as well as its orientation and distance to grid cells with lane boundaries.

[0007] EP 2 679 956 A2 provides a method for identifying discrepancies in digital map data. The method involves selecting one or more candidate locations as a subset of locations within positional data relating to the movement of a large number of devices over a given time within a given area. Each candidate location is assigned to one or more predefined categories based on the distribution of the movement directions of the devices at each candidate location. The candidate locations are then compared with a database of map data, and locations with potential deviations in the digital map data are identified based on the category assigned to each candidate location.

[0008] The invention is based on the objective of providing a novel method for determining the lane layouts of roads.

[0009] The problem is solved according to the invention by a method which has the features specified in claim 1.

[0010] Advantageous embodiments of the invention are the subject of the dependent claims.

[0011] According to the invention, a method for determining and providing lane profiles of roads using a central computer unit digitally coupled to vehicles in a fleet provides that a geometric map with geometric lane boundaries and a grid with grid cells of a predetermined size is provided on the map. Fleet data from vehicles in the fleet are collected, wherein the fleet data includes position sequences driven by the vehicles in the fleet. Furthermore, the vehicle orientations at the positions of the grid cells are determined from the collected position sequences, and the determined vehicle orientations are discretized.Subsequently, for individual grid cell positions, a histogram is generated for the discretized vehicle orientations determined at the position of the respective grid cell, a map section with a predetermined number of grid cells is selected, and lane profiles on the map section are determined using a learning-based method from the histograms created for the grid cells of the map section and the geometric lane boundaries on the map section.

[0012] Applying this method enables the combined processing of geometric infrastructure and fleet driving behavior. This allows for the fully automated derivation of lane paths, whereas derivation based solely on geometric infrastructure always requires manual rework, and derivation of lane paths based on fleet driving behavior does not achieve full coverage.

[0013] By using geometric information, it can be largely ensured that modeled lane layouts are aligned with a given infrastructure.

[0014] Since the method relies on the use of fleet driving behavior, it takes into account human driving style, i.e., manual operation of vehicles within the fleet. This supports lane modeling in unstructured areas, for example, where lane markings are missing and / or bus lanes are absent.

[0015] The learning-based method essentially automatically accounts for special cases if they are represented in the dataset, i.e., in the fleet data. The method is scalable using the fleet data, which means that the determination of lane paths continuously improves with an increasing number of training examples.

[0016] Furthermore, the learning-based method can directly generate lane geometry. Optionally, an older lane layout can also be incorporated into the learning process, so that only parameters of a lane segment regarding anchor point, orientation, curvature, and / or lane width need to be predicted. This simplifies the training process and largely ensures that geometrically valid lane segments are always generated.

[0017] Abstracting the lane paths, which can be determined from the fleet data of the vehicles, into the grid enables consistent processing of the driving behavior of a vehicle belonging to the fleet with varying numbers of trips.

[0018] A multidimensional grid allows for the modeling of multiple directions of travel, for example on bidirectional roads or in intersection areas.

[0019] A grid created based on different time points enables the detection of changes in fleet driving behavior and thus allows conclusions to be drawn about changes in road infrastructure. This also makes it possible to identify changes that are not detected by signals from vehicle sensors, such as a traffic sign prohibiting a turn.

[0020] Since the processing takes place at the level of the grid cells, a decomposition into sub-problems with regard to the determination of the lane paths and thus a parallelization is possible.

[0021] In one implementation, the determined lane paths are made available to the vehicles in the fleet for retrieval. This provision of the determined lane paths enables optimized automated and autonomous driving of the respective vehicle. The respective lane path is known, so that in automated or autonomous driving mode, for example due to missing lane markings, it is not necessary for a driver to take over driving tasks and move the vehicle manually.

[0022] In a further version, the determined lane paths are provided to the vehicles as data from a digital map and stored in it, so that automated or autonomous driving operation of the vehicle can be significantly improved without the need for a driver to intervene.

[0023] One possible implementation of the process involves feeding the determined lane profiles into a trajectory planning module for autonomous vehicles in the fleet. Based on these profiles, a trajectory is then planned for each autonomous vehicle, enabling optimized autonomous driving operation essentially without interruptions from handing off control to a human driver.

[0024] In a further development of the process, the learning-based method is trained using generated histograms and ground truth maps created for predefined regions. This allows the learning-based method to be trained using real-world data and optimized for determining lane paths based on the transmitted fleet data.

[0025] Exemplary embodiments of the invention are explained in more detail below with reference to drawings.

[0026] This shows: Fig. 1 schematically shows a section of a geometric map with geometric track boundaries and a grid with grid cells, Fig. 2 schematically shows a histogram with frequency distributions generated for each grid cell position, and Fig. 3 schematically shows a plurality of tube-like corridors.

[0027] Corresponding parts are marked with the same reference symbols in all figures.

[0028] In Figure 1 The figure shows a section of a geometric map K with geometric track boundaries S and a grid G ​​with grid cells G1 to G49.

[0029] Figure 2 The grid G ​​shows the grid cells G1 to G49 filled using a data structure and in Figure 3A plurality of identified tube-like corridors F are shown.

[0030] It is generally known that automated and autonomous driving requires a highly accurate digital map, i.e., a road map, where the content of the digital map must be correct and up-to-date. For automated driving applications at Level 2+ or higher according to the SAE J3016 standard, such digital maps with lane-accurate topology modeling are mandatory. In the future, mapping will be carried out using signals from vehicle sensors in a fleet of vehicles, particularly those from a single manufacturer. The fleet of vehicles achieves a relatively large area of ​​coverage, and the vehicles transmit fleet data—information relating to a static infrastructure, such as recognized traffic signs and lane markings—to a central computing unit that is linked to the fleet's data network.

[0031] In addition to this information regarding the static infrastructure as environmental information, the position, and in particular the orientation, of the respective vehicle, determined by satellite navigation, is also transmitted to the central computer unit in order to convert a detection relative to a vehicle coordinate system into a global coordinate system. Furthermore, satellite navigation provides information about the driving behavior of individual vehicles, especially about drivable corridors F.

[0032] The following describes a method for determining and providing lane layouts of roads using a central computer unit that is digitally linked to the vehicles in the fleet. This method is designed in multiple stages.

[0033] First, the geometric map K with the lane boundaries S is generated, and the grid G ​​with grid cells G1 to G49 is plotted. The grid cells G1 to G49 have a predefined size relative to the map K. Specifically, the geometric map K is generated from observations of the vehicles in the fleet using established techniques. Individual steps include single-trace pose optimization, multi-trace alignment, and geometric aggregation.

[0034] The central computer unit collects fleet data from the vehicles in the fleet. This fleet data comprises position sequences driven by the vehicles. The fleet data records the respective orientation of the vehicles, with the position sequences having a predefined duration or length. This duration or length is chosen to be relatively short in order to anonymize the collected fleet data and make it difficult to draw conclusions about the movement profile of a particular vehicle. All orientations relative to the geometric map K are located and recorded to determine the lane-specific positions of the vehicles. In particular, the orientations of the vehicles relative to grid cells G1 to G49 are determined from the collected position sequences.

[0035] The determined vehicle orientations are discretized, meaning they are placed within a predefined grid, for example, a 60° grid. In other words, a data structure for the grid G ​​is generated based on a given geographic position. The dependence of the grid G ​​on the geographic position, that is, in relation to the global coordinate system with the coordinates x, y, z, can be discretized into 1 m x 1 m squares using the predefined grid of the geographic map K.

[0036] Subsequently, a histogram H is generated for each grid cell position, representing the discretized vehicle orientations determined at the respective grid cell G1 to G49. Using grid cells G1 to G49, a frequency distribution depending on the direction of travel is modeled multidimensionally to represent geographical points with multiple directions of travel, such as intersections. The frequency distribution can be discretized, for example, by classifying it into 10° classes.

[0037] A learning-based method, for example Convolutional Neural Network, Transformer Network, Graph Neural Network, uses geometric lane boundaries S, in particular lane markings, curbs, etc., from the previously provided geometric map K and the grid G ​​with the frequency distributions from the fleet data as input information and delivers a lane model as output information.

[0038] The machine learning algorithm is trained using ground truth maps, which depict the actual lane configurations for specific regions. The algorithm learns, based on examples, how to generate the lane model from geometric lane boundaries S (lane markings, curbs, etc.) and a histogram H.

[0039] Figure 1Figure 1 shows an example section of a grid G ​​with a superposition of the lane boundaries S. Modes of distribution over the direction of travel are represented by arrows P.

[0040] It is relatively easy to see that in grid cells G1 to G49, which are traversed by vehicles in both directions, two oppositely directed arrows P are shown.

[0041] Grid cells G1 to G49 of a first row Z1 are numbered, where a first hatched corridor F and a second hatched corridor F are the lane profiles generated by the learning-based method.

[0042] In Figure 2 An example histogram H for each grid cell G1 to G7 of grid G ​​is shown. A discretization of the vehicle direction with 60° classes was chosen. In particular, it shows Figure 2a respective histogram H with the values ​​for the respective in Figure 1 The grid cells G1 to G7 shown represent the number of discretized vehicle orientations determined.

[0043] According to a Figure 1 In the grid cell G5 shown, it is evident that a vehicle drove in this cell with an orientation represented by a downward-pointing arrow P. Therefore, a mode (a maximum of a distribution) in the histogram H lies at approximately 200°.

[0044] Grid cells G6 and G7 are traversed with the vehicles oriented in opposite directions, with the mode in histogram H at approximately 20°.

[0045] The respective histogram H based on the grid cells G1 to G7 of the grid G ​​is a comparatively efficient representation of any number of journeys of the grid cells G1 to G49 of the grid G ​​plotted on the geographic map K and represents a fleet driving behavior of the vehicles in a uniform form.

[0046] Using such a process of abstraction, the histogram H serves as a consistent input for the subsequent learning-based procedure, which derives the lane paths by adding a geometric environment.

[0047] Such a uniform representation of a variable number of journeys can also be used to detect changes in fleet driving behavior relatively early in the mapping development process. This allows for the recording of changes in fleet driving behavior that are not detected by signals from vehicle sensors, such as a specific traffic sign that prohibits a turning maneuver.

[0048] The lane alignments generated by the learning-based method are geometrically tube-like corridors F. Therefore, an existing, older lane alignment can optionally be taken into account by the learning-based method, so that the method only needs to predict parameters of a lane segment, such as anchor point, orientation, curvature, and lane width. This simplifies the training process and ensures that geometrically valid lane alignments are always generated. Examples of such tube-like corridors F, which arise from a suitable selection of input parameters for the learning-based method, are shown in Figure 3 shown.

[0049] Furthermore, the procedure provides that the determined lane profiles in the form of the tube-like corridors F are made available to the vehicles of the vehicle fleet as data in the digital maps.

[0050] In particular, the determined lane profiles are fed into a trajectory planning module of an automated or autonomous vehicle, so that a trajectory of the vehicle can be planned according to the existing lane profiles.

Claims

1. Method for determining and providing lane courses of roads by means of a central computing unit which is data-linked to vehicles in a vehicle fleet, characterized by the steps of: - providing a geometric map (K) having geometric lane boundaries (S) and applying a grid (G) having grid cells (G1 to G49) of a predefined size to the map (K); - collecting fleet data from vehicles in the vehicle fleet, the fleet data comprising position sequences driven by the vehicles in the vehicle fleet; - determining vehicle orientations of the vehicles at positions of the grid cells (G1 to G49) from the collected position sequences; - discretizing the determined vehicle orientations; - generating, for each individual grid cell position, a histogram (H) for the discretized vehicle orientations determined at the position of the relevant grid cell (G1 to G49); - selecting a map section having a predefined set of grid cells (G1 to G49); and - determining the lane courses on the map section from the histograms (H) created for the grid cells (G1 to G49) of the map section and the geometric lane boundaries (S) on the map section by means of a learning-based method.

2. Method according to claim 1, characterized in that the determined lane courses are provided to the vehicles in the vehicle fleet for retrieval.

3. Method according to claim 1 or claim 2, characterized in that the determined lane courses are provided to the vehicles as data in a digital map.

4. Method according to any of the preceding claims, characterized in that the determined lane courses are supplied to a trajectory planning module of autonomously driving vehicles in the vehicle fleet.

5. Method according to any of the preceding claims, characterized in that training of the learning-based method is carried out by means of created histograms (H) and by means of ground truth maps generated for predefined regions.